<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Shuai Chen on Medium]]></title>
        <description><![CDATA[Stories by Shuai Chen on Medium]]></description>
        <link>https://medium.com/@schen583?source=rss-e5271e521780------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*dEwy4-Q1X8JDCp02LD80QQ.jpeg</url>
            <title>Stories by Shuai Chen on Medium</title>
            <link>https://medium.com/@schen583?source=rss-e5271e521780------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 16:23:07 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@schen583/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[China’s Emerging National Standard for L3 & L4 Autonomous Driving]]></title>
            <link>https://schen583.medium.com/chinas-emerging-national-standard-for-l3-l4-autonomous-driving-ea8cd8fe1bba?source=rss-e5271e521780------2</link>
            <guid isPermaLink="false">https://medium.com/p/ea8cd8fe1bba</guid>
            <category><![CDATA[china]]></category>
            <category><![CDATA[autonomous-vehicles]]></category>
            <category><![CDATA[system-engineering]]></category>
            <category><![CDATA[technology-policy]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <dc:creator><![CDATA[Shuai Chen]]></dc:creator>
            <pubDate>Tue, 30 Jun 2026 06:08:05 GMT</pubDate>
            <atom:updated>2026-06-30T06:08:05.987Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XSqGIW3-A87HsEowgSjLZw.jpeg" /><figcaption>System redundancy architecture illustrating multi‑layer fail‑operational and fail‑safe design principles. Generated with AI.</figcaption></figure><p><em>Following my previous post on L2 ADAS regulations in China, I continue the discussion by examining the draft L3 and L4 autonomous‑driving regulations circulated earlier this June.</em></p><p>China’s mandatory national standard for L3 and L4 autonomous driving has now advanced to the “draft for approval” stage, transforming the 2024 technical guideline (GB/T 44721‑2024) into a full safety‑case‑based, type‑approval regulation with scenario‑level requirements, SMS governance, simulation validation, and explicit compliance‑verification procedures. This marks the country’s most comprehensive ADS safety regulation to date and is expected to take effect on July 1, 2027. The standard also unifies terminology by formally naming both L3 and L4 systems as Automated Driving Systems (ADS).</p><h3><strong>Core Technical Requirements for ADS</strong></h3><h4>Baseline Requirements</h4><p>The draft standard defines the safety level of an ADS as at least equivalent to a qualified and attentive human driver, performing the driving task without imposing unreasonable risk on users or other road users (ORU). It adopts the globally recognized concept of behavioral competency, requiring that an ADS — when operating within its operational design condition (ODC) — exhibit driving behaviors that are predictable, appropriate, and demonstrably competent.</p><h4>L3 vs. L4 Functional Capability Thresholds</h4><p>The standard differentiates L3 and L4 requirements particularly for non‑expressway roads, including urban environments:</p><ul><li>L3 ADS shall support lane‑keeping cruise, lane‑change control, and intersection traversal (excluding roundabouts).</li><li>L4 ADS shall additionally handle obstacle‑bypassing with partial lane entry, reversing, and U‑turn maneuvers.</li></ul><p>These distinctions reflect the higher autonomy expectations for L4 systems, which operate without a fallback‑ready human driver.</p><h4>Requirements Across Four Scenario Categories</h4><p>The standard structures Dynamic Driving Task (DDT) execution into four scenario categories, each with distinct behavioral expectations.</p><p>1. Nominal Scenarios</p><p>Under nominal conditions, the ADS shall:</p><ul><li>Avoid collisions and prevent contact with safety‑relevant objects.</li><li>Anticipate risks, adjust speed according to risk levels, and maintain safe distance from ORUs.</li><li>Handle areas that cannot be fully sensed (e.g., blind spots).</li><li>Interact safely with ORUs using appropriate signaling (turn indicators, brake lights).</li><li>Respond correctly to priority vehicles (police, fire, ambulance) and on‑site traffic police instructions.</li></ul><p>A key distinction:</p><ul><li>L3 may either yield or perform a fallback response.</li><li>L4 must yield and obey police commands.</li></ul><p>2. Risk Scenarios</p><p>When a collision is unavoidable, the ADS shall:</p><ul><li>Minimize injury or damage.</li><li>Bring the vehicle to a stop after the collision.</li></ul><p>Whether the collision could have been avoided is assessed against the performance of a competent human driver.</p><p>3. Failure Scenarios</p><p>The ADS shall:</p><ul><li>Detect faults that affect DDT execution.</li><li>Execute a fallback response when safe DDT performance is no longer possible.</li><li>Disable activation of the affected function until the fault is resolved.</li></ul><p>4. Out‑of‑ODC Scenarios</p><p>For foreseeable departures from the ODC, the ADS shall anticipate them.</p><ul><li>If not yet activated, the ADS shall prevent activation when encountering conditions outside its ODC.</li><li>If already activated, it shall execute a fallback response upon encountering such conditions.</li></ul><h4>Minimum Risk Maneuver (MRM)</h4><p>When the ADS can no longer safely continue DDT execution, it shall perform a Minimum Risk Maneuver (MRM) to reach a Minimum Risk Condition (MRC).</p><p>Both L3 and L4 systems shall:</p><ul><li>Be capable of lane changes during MRM.</li><li>Minimize risk throughout the maneuver.</li><li>Pull over to a roadside position that does not obstruct traffic, or — if movement is impossible — stop safely within the current lane.</li><li>Use hazard warning signals appropriately.</li></ul><p>A specific L3 requirement:</p><ul><li>After reaching MRC, L3 ADS cannot be reactivated unless the power system is restarted.</li></ul><h3><strong>Human–Machine Interaction (HMI)</strong></h3><h4>General HMI Requirements</h4><p>The ADS shall not auto‑activate after power‑on. It shall clearly communicate its status, faults, and state transitions to users through multi‑modal cues — including visual, auditory, and haptic signals. These cues ensure that users can reliably understand when the ADS is available, active, degraded, or transitioning into fallback or MRM states.</p><h4>L3 Fallback‑User Monitoring Scheme</h4><p>The draft standard introduces a three‑layer monitoring scheme to ensure the fallback‑ready user remains “in a ready state” during L3 operation:</p><ul><li>Seatbelt monitoring — If the seatbelt is unfastened, the system shall issue a takeover request.</li><li>Seating‑position monitoring — If the fallback‑ready user is not seated in the driver’s seat for more than 1 second, a takeover request shall be issued.</li><li>DDT‑capability monitoring — Within a period not exceeding 30 seconds, the system shall confirm — using at least two independent indicators (e.g., HMI interactions, eye movements, head movements, body movements, voice, physiological features) — that the fallback‑ready user is capable of taking over.</li></ul><p>— If fewer than two indicators are available, or if the system determines the user is not capable, an insufficient‑capability warning shall be issued immediately.</p><p>— If capability is not restored within 15 seconds, the warning shall escalate to a takeover request.</p><p>This three‑layer scheme is one of the most stringent L3 human‑state monitoring frameworks globally.</p><h4>L3 Takeover Request Timing and Escalation</h4><p>The standard specifies detailed timing requirements for takeover requests in L3 ADS:</p><ul><li>Planned takeover events shall be issued early enough that, even without user intervention, an MRM can safely bring the vehicle to a stop before the planned event occurs.</li><li>Unplanned takeover events shall trigger a takeover request immediately upon detection.</li><li>A takeover request shall escalate within 4 seconds after issuance and remain in the escalated state.</li><li>From initial issuance until termination due to MRM, the takeover‑request duration shall be no less than 10 seconds, ensuring the fallback‑ready user has sufficient time to respond.</li><li>In cases of severe failures, the system shall proceed directly to MRM without issuing a takeover request.</li><li>During takeover requests and MRM, the ADS shall provide continuous status indications to users.</li></ul><p>These rules create a predictable, auditable takeover‑request sequence for L3 systems.</p><h4>Duty‑to‑Inform + Duty‑to‑Acknowledge</h4><p>ADS safety includes both technical obligations and information‑disclosure obligations.</p><p>OEMs shall provide publicly accessible ADS usage instructions, covering:</p><ul><li>System capabilities and limitations</li><li>Activation, exit, takeover, intervention, and MRM procedures</li><li>Role transitions between user and ADS</li><li>Risks associated with non‑driving activities</li><li>Emergency handling procedures</li></ul><p>Each time the power system is restarted, the ADS shall confirm that the user has completed the required ADS training — via training confirmation, biometric verification, or account login.</p><p>This confirmation:</p><ul><li>Requires a long‑press or two deliberate actions</li><li>Shall occur only when the vehicle is stationary</li><li>Shall be re‑confirmed after any software update that changes the instructions</li><li>Shall be re‑confirmed if no confirmation has been provided within 30 days</li></ul><p>This mechanism applies to both L3 and L4, though its practical burden is heavier for L3.</p><h4>L4 Passenger Safety and Remote Assistance</h4><p>For L4 vehicles, the standard adds requirements to protect passengers:</p><ul><li>The system shall provide safety‑relevant information to passengers.</li><li>It shall respond to risks such as unfastened seatbelts or improper seating, according to the Safety Case strategy.</li><li>It shall prevent passengers from accessing or manipulating driving‑control elements.</li></ul><p>If the L4 ADS includes a remote assistance function:</p><ul><li>Remote assistance shall not be used to execute the DDT.</li><li>It may only be triggered under conditions defined in the Safety Case, such as:</li></ul><p>— Collision</p><p>— Vehicle stuck</p><p>— ADS failure</p><p>— Vehicle failure</p><p>Remote assistance is strictly a support mechanism, not a driving function.</p><h3><strong>Safety Maintenance Specification (SMS)</strong></h3><h4>Definition and Scope</h4><p>According to definition 3.27, the Safety Management System (SMS) is a systematic safety‑management framework integrating organizational, personnel, and technical components. Manufacturers shall establish and implement an SMS that covers:</p><ul><li>Safety policy</li><li>Risk management</li><li>Safety assurance</li><li>Continuous safety improvement</li><li>Design &amp; development safety management</li><li>Production safety management</li><li>Post‑deployment safety management</li></ul><p>The standard then specifies detailed expectations for several of these elements.</p><h4>Lifecycle Safety Management Requirements</h4><ul><li>Risk management shall span the entire ADS lifecycle — development, production, and post‑deployment — and shall address risks arising from all participating parties, including suppliers and service providers.</li><li>Safety assurance shall include periodic independent internal audits, external assessments, supplier SMS capability evaluations, and the establishment of KPIs to measure SMS effectiveness.</li><li>Design &amp; development safety management shall integrate functional safety, SOTIF, and human‑factors engineering, supported by continuous evaluation using FMEA, FTA, STPA, and similar methodologies.</li><li>Post‑deployment safety management requires manufacturers to maintain safety monitoring and safety‑management capabilities throughout the operational life of the ADS.</li></ul><h4>Verification and Validation Requirements</h4><p>The standard also specifies requirements for test conditions, with particular emphasis on simulation testing. Manufacturers shall:</p><ul><li>Demonstrate verification and validation of the simulation toolchain.</li><li>Quantify uncertainties introduced by data quality.</li><li>Manage variance and ensure reproducibility of stochastic models.</li><li>Justify personnel competence involved in simulation development and evaluation.</li><li>Conduct release management and criticality analysis for simulation assets.</li></ul><p>These requirements ensure that simulation‑based evidence used in the Safety Case is trustworthy and auditable.</p><h4>Proving‑Ground and Road Testing Requirements</h4><p>For physical testing, the standard requires that:</p><ul><li>Proving‑ground facilities and road‑test routes shall align with the ADS’s intended use.</li><li>Test equipment shall undergo periodic calibration.</li><li>The ADS shall be exposed to sufficiently probable interactions with diverse ORUs and atypical conditions, ensuring that testing reflects realistic operational complexity.</li></ul><p>These provisions ensure that real‑world testing complements simulation and covers the breadth of operational scenarios relevant to the ADS.</p><h3><strong>Safety Case</strong></h3><h4>Safety Case Framework</h4><p>Appendix D provides a complete Safety Case framework built around claims, arguments, and evidence, consistent with global practice in safety‑critical industries.</p><p>The standard requires that:</p><ul><li>Every claim shall be supported by at least one argument.</li><li>Every argument shall be supported by at least one item of evidence.</li><li>All claims, arguments, and evidence shall have unique identifiers and maintain full traceability from the originating requirement to the supporting evidence.</li><li>For each technical requirement, the Safety Case shall include at least one corresponding claim demonstrating how the requirement is satisfied.</li></ul><p>This establishes a structured, auditable chain from requirements → claims → arguments → evidence.</p><h4>Residual‑Risk Acceptance Criteria</h4><p>Manufacturers shall define ADS residual‑risk acceptance criteria using:</p><ul><li>Human‑driving accident data</li><li>Driving‑assistance accident data</li><li>Automated‑driving accident data</li><li>Industry benchmarks and advanced technologies</li></ul><p>Residual‑risk criteria form the quantitative and qualitative basis for evaluating whether the ADS is acceptably safe for its intended use.</p><h4>Functional Safety and SOTIF (Performance‑Insufficiency Risks)</h4><p>The Safety Case shall document both functional safety and SOTIF measures.</p><p>SOTIF addresses risks arising from performance insufficiencies rather than system faults, triggered by:</p><ul><li>Environmental factors (rain, snow, fog, glare, dust)</li><li>Roadway conditions (curves, ramps, tunnels, missing lane markings, construction zones)</li><li>Object characteristics (non‑standard vehicle shapes, low‑contrast objects, umbrella‑holding pedestrians)</li><li>Multi‑object interactions (occlusion, crossing, cut‑ins)</li><li>User misuse</li></ul><p>For each category, manufacturers shall define residual risk for both known and unknown hazardous scenarios, supported by combined evidence from:</p><ul><li>Simulation</li><li>Proving‑ground tests</li><li>Road tests</li></ul><p>Safety analysis shall integrate both top‑down and bottom‑up approaches, employing methodologies such as FMEA, FTA, HAZOP, and STPA.</p><h4>Independent Review Prior to Type-Approval</h4><p>Before type-approval, the manufacturer shall have the Safety Case reviewed by internal auditors who are organizationally independent from the development team.</p><p>All findings and corrective actions shall be documented, ensuring that the Safety Case is:</p><ul><li>Independently evaluated</li><li>Free from development‑team bias</li><li>Supported by traceable corrective‑action records</li></ul><p>This independent review is a critical regulatory safeguard ensuring the credibility of the Safety Case.</p><h3><strong>Validation, Type-Approval, and Implementation</strong></h3><h4>Overall Approval Process</h4><p>The approval process consists of three major components: SMS audit, Safety Case audit, and validation testing. Together, these ensure that ADS safety is demonstrable, reproducible, and sustainably maintained throughout the system’s lifecycle.</p><h4>SMS Audit</h4><p>Inspectors shall verify that the manufacturer’s SMS complies with Chapter 6 of the standard.</p><p>This includes confirming that safety‑related processes across the entire ADS lifecycle — development, production, and deployment — are robust, auditable, and consistently implemented.</p><p>The SMS audit ensures that organizational safety governance, personnel competence, and technical processes are sufficiently mature to support ADS safety claims.</p><h4>Safety Case Audit</h4><p>Inspectors shall evaluate the completeness and robustness of the Safety Case submitted by the manufacturer, following Appendix D.</p><ul><li>Completeness — whether claims, arguments, and evidence are sufficient, traceable, and correctly linked.</li><li>Robustness — whether residual risks fall below the unreasonable‑risk threshold, whether evidence is reproducible, and whether foreseeable conditions inside and outside the ODC are reasonably covered.</li></ul><p>This audit ensures the Safety Case is not merely documentation but a defensible, evidence‑based demonstration of ADS safety.</p><h4>Validation Testing</h4><p>Inspectors shall verify ADS performance using:</p><ul><li>Proving‑ground tests (GB/T 41798)</li><li>Road tests (GB/T 44719)</li><li>Simulation tests (GB/T 47025)</li></ul><p>These tests confirm that the ADS behaves safely under real and simulated conditions and that the Safety Case is supported by actual test results, not just theoretical arguments.</p><h4>Type‑Approval Rules</h4><p>Same‑Type Definition</p><p>An ADS is considered the same-type when the following elements are identical:</p><ul><li>Vehicle manufacturer</li><li>Perception and positioning system</li><li>ECU hardware</li><li>Software architecture</li><li>ODC configuration</li><li>Automated‑driving functions</li><li>MRM strategy</li></ul><p>The type‑approval report remains valid for three years.</p><p>Software‑version differences are permissible only if they do not affect ADS performance.</p><p>Supplementary Inspection for Limited Changes</p><p>If the only changes relate to:</p><ul><li>Road type</li><li>Time of day</li><li>Maximum operating speed</li><li>Weather conditions</li><li>Lane type (same / reduced / lowered)</li></ul><p>then only supplementary inspection of the modified parameters is required, rather than full re‑approval.</p><h4>Implementation Timeline</h4><ul><li>For newly applied type‑approval vehicle models, the requirements take effect from the date of implementation.</li><li>For vehicle models already holding type-approval, the requirements take effect from the 13th month following the implementation date.</li></ul><p>This phased approach ensures regulatory continuity while giving manufacturers time to adapt.</p><h3><strong>System Redundancy</strong></h3><h4>Overall Redundancy Requirements</h4><p>The standard establishes redundancy requirements across functional safety, MRM capability, sensor/ECU architecture, remote assistance, and Safety Case evidence.</p><p>The goal is to ensure the ADS can either continue safe operation or safely degrade when faults occur.</p><h4>No Single‑Point‑Failure (SPF) Requirements</h4><p>Clause D.2.4.2.5 explicitly states that the ADS shall not lose steering, braking, or parking capability due to any single‑point failure (SPF) in the electrical/electronic (E/E) system.</p><p>This implies redundancy for:</p><ul><li>Steering actuation</li><li>Braking actuation</li><li>Parking brake control</li><li>Power supply</li><li>Communication paths</li></ul><p>These are foundational requirements ensuring that critical actuation functions remain available even under fault conditions.</p><h4>Fault‑Detection Coverage (FDC) and Diagnostic Redundancy</h4><p>The first note of D.2.4.2.5 links redundancy to fault‑detection coverage (FDC).</p><p>FDC shall be justified based on:</p><ul><li>Fault likelihood</li><li>Fault severity</li><li>Whether user takeover is possible</li><li>Whether MRM can still be executed</li></ul><p>This implies diagnostic redundancy, including:</p><ul><li>Dual sensors for cross‑checking</li><li>Dual compute paths for monitoring</li><li>Redundant communication channels for heartbeat monitoring</li><li>Independent fault‑detection mechanisms</li></ul><p>These measures ensure that faults are detected reliably and early enough to trigger safe fallback behavior.</p><h4>MRM Robustness Under Failures</h4><p>The second note of D.2.4.2.5 emphasizes that MRM shall remain robust under failures, not only in nominal conditions.</p><p>Later, D.2.4.2.8 requires the Safety Case to validate that residual risk remains acceptable even when MRM is executed under severe faults.</p><p>This ensures that MRM is treated as a safety‑critical capability, not a convenience feature.</p><h4>Redundancy Design Principles</h4><p>Clause D.2.4.2.4 defines four redundancy design principles:</p><ul><li>Internal degraded‑mode operation</li><li>Independent system</li><li>Heterogeneous system</li><li>Functional degradation</li></ul><p>Manufacturers may apply one or more of these principles depending on system architecture and safety goals.</p><h4>L3 Takeover‑Request Redundancy</h4><p>Clause D.2.4.2.7 mandates that an L3 ADS shall not lose its ability to issue takeover requests due to any single E/E failure.</p><p>This requirement ensures that the fallback‑ready user can always be alerted when ADS performance is compromised.</p><h4>Remote‑Assistance Communication Redundancy (Primarily L4)</h4><p>Clause D.2.4.2.10 requires that the ADS — primarily L4 systems — shall be able to handle communication faults related to remote assistance.</p><p>Remote assistance may be used to support decision‑making in complex or ambiguous situations, so communication paths shall be fault‑tolerant.</p><h4>Safety Case Documentation of Redundancy</h4><p>Clause D.1.3.1(c)(6) requires the Safety Case to explicitly describe all redundant components and redundant connections in the ADS architecture.</p><p>This ensures full traceability and transparency of redundancy design choices.</p><p>China’s move toward formal L3 and L4 regulation marks a clear shift from guidance to enforceable safety requirements. As the standard approaches its 2027 implementation, manufacturers and suppliers will need to align their systems, processes, and evidence with a far more rigorous approval framework — one that will shape the next stage of automated‑driving deployment.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ea8cd8fe1bba" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A First Look at China’s New National Standard for L2 ADAS]]></title>
            <link>https://schen583.medium.com/a-first-look-at-chinas-new-national-standard-for-l2-adas-5fb91a089280?source=rss-e5271e521780------2</link>
            <guid isPermaLink="false">https://medium.com/p/5fb91a089280</guid>
            <category><![CDATA[automotivesafety]]></category>
            <category><![CDATA[ada]]></category>
            <category><![CDATA[simulation]]></category>
            <category><![CDATA[china-tech]]></category>
            <category><![CDATA[autonomous-vehicles]]></category>
            <dc:creator><![CDATA[Shuai Chen]]></dc:creator>
            <pubDate>Sun, 31 May 2026 14:23:42 GMT</pubDate>
            <atom:updated>2026-05-31T14:23:42.923Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TxbJg4g1mwkR4W0d8HB1hQ.png" /></figure><p>China’s first mandatory national standard governing L2 driver assistance has advanced to the “draft for approval” stage, indicating that the technical content is essentially finalized. Although conceptually aligned with UN R171, the standard reflects China’s specific regulatory priorities, particularly around highway and urban NOA. Implementation is expected in January 2027.</p><p>The document defines safety requirements, test methods and approval criteria for the<strong> </strong>Combined Driver Assistance System (CDAS), a system that continuously controls both lateral and longitudinal motion within its ODD. It applies to vehicle categories of M and N.</p><h3><strong>Three Categories of CDAS Systems</strong></h3><p><strong>Basic Single-Lane CDAS</strong></p><ul><li>Operates only on A‑type roads (highways/expressways).</li><li>Maintains vehicle control within a single lane.</li><li>RMF (Risk Mitigation Function) is not required.</li></ul><p><strong>Basic Multi-Lane CDAS</strong></p><ul><li>Also limited to A‑type roads.</li><li>Supports driver‑triggered lane changes only.</li><li>RMF is required.</li></ul><p><strong>Advanced (Navigation) CDAS</strong></p><ul><li>Operates on A‑type and/or B‑type roads.</li><li>Supports lane changes (driver initiated, driver confirmed, system initiated), obstacle avoidance, intersection handling, roundabouts, and other non‑lane‑cruise functions.</li><li>RMF is required.</li><li>Highway NOA and Urban NOA fall under this category.</li></ul><h3><strong>Preventing Driver Misuse: Continuous Monitoring, Alert Escalation, and System Disablement</strong></h3><p>The framework forces the driver to stay engaged by requiring two simultaneous monitoring channels: hands-off detection and eyes-off detection. Both shall be active whenever CDAS is launched and vehicle speed is ≥10 km/h. If either disengagement behavior is detected, the system shall follow a non‑bypassable escalation chain with hard timing limits — illustrated in Figure 1 — that may lead to an RMF‑controlled stop.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/490/1*0n8EnGsIqfdp_pvpiGX17w.png" /></figure><p>The standard also initiates system disablement penalties for the first time. If driver disengagement triggers any of the following events during the current ignition cycle, CDAS shall be disabled for no less than 30 minutes:</p><ul><li>One RMF activation.</li><li>If the system does not support RMF, it continuously issues an escalated HOR or DCA for 10 seconds.</li><li>The system issues two DCA alerts escalated from EOR.</li><li>Within any 30‑minute period, the system issues three escalated HOR or EOR alerts.</li></ul><h3><strong>Mandatory Driver Education and Confirmation</strong></h3><p>OEMs shall provide system-usage instructions to drivers through channels such as official websites, in‑vehicle display terminals, or smart mobile devices. The instructions shall clearly explain the system category, driver responsibilities, system capabilities and limitations, ODD boundaries, designed operating-speed range of the system, driver-disengagement detection methods and escalation sequence, among other required content.</p><p>During each vehicle ignition cycle, the system shall confirm whether the driver has completed system usage training through at least one of the following methods:</p><ul><li>Conducting system usage training for the driver and confirming completion.</li><li>Biometric identification.</li><li>Account login while the vehicle is stationary.</li></ul><p>The vehicle shall provide a method for the driver to confirm that they have read and understood the system-usage instructions. This confirmation method shall meet the following requirements:</p><ul><li>Confirmation is only allowed when the vehicle is stationary.</li><li>The confirmation method shall require a long‑press/hold or at least two deliberate actions (e.g., double‑click, click‑and‑confirm).</li><li>If the instructions change due to a software update, the confirmation method shall be presented before the next system activation.</li><li>If the system detects that the driver has not confirmed within 30 days, the confirmation method shall be presented before the next system activation.</li></ul><h3><strong>Hard KPIs for Functional Requirements</strong></h3><p>The framework establishes a series of measurable, quantitatively defined KPIs governing CDAS functional behavior. Exhibit 1 summarizes the key system thresholds and performance requirements.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/429/1*Wp9HwiXL3NjgBRm0rzz9xg.png" /></figure><h3><strong>Corner Cases Included in the Closed‑Course Test Method</strong></h3><p>The standard contains the most extensive chapter dedicated to closed‑course test procedures required to verify CDAS safety performance. A selection of representative corner‑case scenarios included in the test method is illustrated below.</p><p><strong>Clause 9.4.7 — Construction‑Zone Detection and Response Capability Test</strong></p><p>Per Figure 2, the test road shall be a long, straight, two‑way, two‑lane roadway. The test vehicle shall travel one lane away from the road centerline. Ahead of the test vehicle, a construction sign and a detour sign shall be installed. Rows of traffic cones shall be used to close the lane. On the oncoming‑traffic side, the inner lane shall be closed using movable barriers, which shall extend to the outer‑lane boundary of the oncoming inner lane.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/832/1*Mk8Bz4w7Ncjs7upi85fHug.png" /><figcaption>Figure 2. Construction‑Zone Detection and Response Capability Test</figcaption></figure><p>The test vehicle shall drive in the right‑hand lane at Vsmaxset‑con (the maximum set speed specified for the construction‑zone scenario). When the distance to the first speed‑limit sign reaches at least 200 m, the system shall be activated and the test shall begin.</p><p>For M1‑category vehicles, the test is considered a fail if the time‑to‑collision (TTC) reaches 1.0 s to the traffic cones and the vehicle still performs no deceleration or steering action.</p><p>If the system supports activation at night, this scenario shall also be executed under nighttime conditions.</p><p><strong>Clause 9.4.6.5 — Straight‑Road Overturned‑Vehicle Test</strong></p><p>Per Figure 3, the test road shall be a one‑way, two‑lane road with a dashed line separating the lanes. In the right‑hand lane ahead of the test vehicle, a stationary overturned‑vehicle target shall be placed. The roof of the overturned vehicle shall face the direction from which the test vehicle approaches, and the front end of the overturned vehicle shall face the left lane.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/832/1*0A0an5d5IsV5Bjgg60fBlQ.png" /><figcaption>Figure 3. Straight‑Road Overturned‑Vehicle Test</figcaption></figure><p>The test vehicle shall drive in the lane containing the overturned‑vehicle target at Vsmaxset. When the distance to the overturned‑vehicle target reaches at least 200 m, the system shall be activated and the test shall begin.</p><p>For M1‑category vehicles, the test is considered a fail if the TTC reaches 1.0 s to the overturned‑vehicle target and the vehicle still performs no deceleration or steering action.</p><p>If the system supports activation at night, this test scenario shall also be executed under nighttime conditions.</p><p><strong>Clause 9.4.6.6 &amp; 9.4.6.7 — Target-Vehicle Cut-In &amp; Cut-Out Test</strong></p><p>Per Figure 4, the test vehicle shall drive in its lane at Vsmaxset. A target vehicle in the left adjacent lane shall travel in the same direction at (Vsmaxset − 50) km/h. For M1‑category vehicles, the target vehicle shall cut into the test vehicle’s lane with a maximum lateral acceleration of 1.5 m/s², a lane‑change duration of 3 s, and a lateral offset of 3.5 m, resulting in a predicted collision time of 1.5 s.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/832/1*bYgcnvYRn8uYTa9WpVWwoA.png" /><figcaption>Figure 4. Target-Vehicle Cut-In Test</figcaption></figure><p>The system shall detect the cut‑in, respond appropriately, and avoid collision before the predicted collision time.</p><p>Per Figure 5, the test vehicle shall drive at Vsmaxset, steadily following a moving target vehicle ahead at (Vsmaxset − 20) km/h. When the predicted collision time between the moving target and a stationary target vehicle ahead reaches 3 s, the moving target shall initiate a lane‑change maneuver to leave the lane.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/832/1*ze7wJtmZH8GbRAMIdsN-hw.png" /><figcaption>Figure 5. Target-Vehicle Cut-Out Test</figcaption></figure><p>The cut‑out scenario represents the case in which the lead vehicle suddenly departs, requiring the test vehicle to immediately detect and respond to a previously hidden stationary object — one of the most challenging situations for AEB systems.</p><p>For M1‑category vehicles, the test vehicle shall have already initiated a collision‑avoidance action when TTC = 1.0 s.</p><p><strong>Other China‑Specific Test Scenarios</strong></p><p>Clause 9.4.8 requires the system to detect speed, lighting, rain, and fog ODD boundaries and shall issue a DCA before the vehicle exits the valid ODD.</p><p>Clause 9.4.6.2 verifies that during a driver‑triggered lane change, the system shall detect a lateral‑conflict vehicle straddling the lane line and respond in time to prevent a side collision.</p><p>Clause 9.5.2 requires the system to safely enter, circulate within, and exit a roundabout featuring a central island ≥ 20 m in diameter and at least three exits, and the system shall exit via the specified exit without error.</p><p>Clause 9.5.3 requires the system to correctly recognize traffic lights at a four‑way intersection and shall strictly obey red‑stop / green‑go behavior.</p><p>Clause 9.5.4 requires the system to safely execute straight‑through and turning maneuvers at a standard intersection, which is a core capability of advanced CDAS operating on urban roads.</p><h3><strong>DSSCDA: Ensuring Safety Traceability and Regulatory Compliance</strong></h3><p>Appendix F defines two types of DSSCDA (Data Storage System for Combined Driver Assistance) system, depending on the vehicle category (M or N) and its sub-category. It covers three classes of recorded events: timestamp events, collision events, and events involving collision risk.</p><p>The scope of data that the DSSCDA shall record includes:</p><ul><li>Vehicle state and dynamics — vehicle speed, acceleration, yaw rate, steering angle, and steering‑angle rate.</li><li>CDAS operational information — system state, activation category, currently active function, and the timestamps of all prompts and warning signals.</li><li>Driving‑environment information — key detected targets, road‑environment information, and external image or video data (sampling rate of ≥ 4 Hz or 4 fps), all carrying timestamps with hundred‑millisecond‑level resolution.</li><li>Driver operation and driver‑state information — steering‑wheel torque, brake/accelerator pedal positions, seat‑belt status, and driver‑disengagement signals.</li></ul><p>As an automotive hardware component, the DSSCDA shall maintain data integrity under vibration, temperature, and EMC conditions, and provide power‑loss robustness, ensuring reliable data retention throughout CDAS activation.</p><p>Furthermore, vehicle manufacturers shall provide data‑reading methods, access to data‑reading tools, and a publicly accessible data‑reading guidance manual, ensuring that third parties can independently perform data extraction.</p><h3><strong>Simulation Testing Adopted for CDAS Validation</strong></h3><p>Safety validation of CDAS systems involves an enormous number of scenarios, and relying solely on real-vehicle testing is neither realistic nor economically viable. However, the credibility of simulation has long lacked a unified specification, with each company applying its own methods.</p><p>Appendix B establishes a rigorous, end‑to‑end framework to ensure that the simulation tools, models, data, and processes used in CDAS safety validation are credible, traceable, verifiable, and representative of real‑world behavior. This framework forms the foundation for trustworthy simulation‑based validation and supports consistent regulatory compliance across the industry.</p><h3><strong>Implementation and Transitional Arrangements</strong></h3><ul><li>For new type-approval applications, vehicles must comply with all requirements — except a small number of deferred clauses — starting from the implementation date. Beginning in the 13th month after implementation, vehicles must comply with all requirements in full.</li><li>For in‑production vehicle types that have already obtained type approval, manufacturers must comply with all relevant requirements, except certain data‑recording sub‑items, starting from the 13th month after implementation. Beginning in the 25th month, these vehicles must comply with all requirements in full.</li></ul><p>As the draft moves toward formal publication, it sets the stage for a more consistent and safety‑anchored L2 landscape in China. By pairing clear definitions with measurable performance requirements and modern validation tools, the standard provides a practical framework for both industry implementation and regulatory oversight. Once it enters into force in 2027, it will shape how L2 systems are developed, deployed, and trusted across the market.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5fb91a089280" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[China ADAS SoC Innovators Overview]]></title>
            <link>https://schen583.medium.com/china-adas-soc-innovators-overview-413e54972d44?source=rss-e5271e521780------2</link>
            <guid isPermaLink="false">https://medium.com/p/413e54972d44</guid>
            <category><![CDATA[semiconductors]]></category>
            <category><![CDATA[automotive]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[china]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Shuai Chen]]></dc:creator>
            <pubDate>Tue, 28 Apr 2026 08:00:43 GMT</pubDate>
            <atom:updated>2026-04-28T08:00:43.577Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*2ot5051IJbBCqf1nSqyUow.png" /></figure><p>In my January <a href="https://medium.com/@schen583/china-automakers-and-tier1s-in-house-adas-chips-to-challenge-nvidia-thor-orin-4debcc9bd3c3">post</a>, I reviewed a list of Chinese OEMs and Tier1s designing ADAS chips in-house, the scales of which are still limited now. Notably, a strong force to innovate on ADAS SoCs are independent semiconductor companies. Now we look into the domestic vendors on their product offerings and market strategies over the years to understand how each player strives to differentiate on a highly competitive market segment.</p><h3><strong>Horizon Robotics</strong></h3><p>Since established in 2015, Horizon Robotics’ product strategy has shifted. Given the founder Kai Yu’s academic background in AI and industrial practice at Baidu driving the integration of deep learning into search engine as well as initiating the autonomous driving project, Horizon was aiming at hardware and software integrated solutions for edge AI, envisioned by Yu as critical to the AI boom.</p><p>As an attempt, the company launched its first-generation chips, Sunrise 1 for AIoT applications and Journey 1 for intelligent driving in 2017. In 2019, it made a bold move to suspend the AIoT business and concentrate on the automotive market. Horizon launched Journey 2 in late 2019 and Journey 3 in 2020, both targeting smart front camera systems enabling regulatory L2 ADAS functions. To differentiate from Mobileye, a market dominator at the time, Horizon offered perception stack in white box allowing customizations with intensive local support. In a way, it helped to secured initial projects from Changan and Li Auto, followed by many other OEMs.</p><p>With Journey 5’s launch in 2021, Horizon started to embrace a more open ecosystem strategy by teaming up with several tier1s. Basically, it took a step back as tier2 by offering mainly chips and development kit. In the following years, Journey 5 has been widely adopted in L2+ systems with highway NOA functions. However, the chip was designed for CNN-based perception algorithm with poor adaptability to transformer-dominated framework that became popular right after Journey 5’s launch. Taking 2023 as an example, it was far behind Orin-X in terms of shipment, despite being more affordable.</p><p>Lesson learned when it came to Journey 6. It was launched as a product family in 2024 with six variants that can power low, mid and high-level ADAS systems. While it continues to cultivate its growing ecosystem partners, Horizon has picked back up the tier1 role to deliver full stack solutions, later revealed as HSD (Horizon SuperDrive). It claimed that it can design better chips by getting the hands dirty.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yfPvMbTPk4QGH4NPjt9zDg.jpeg" /><figcaption>Horizon Journey 6 series</figcaption></figure><p>A parallel story was the spin-off of its AIoT business line as D-Robotics in 2024. Despite the changes Horizon has made over the past 10 years, it seems like the company is still on its mission to create integrated solutions for edge AI that include automotive and robotics.</p><h3><strong>Black Sesame</strong></h3><p>Black Sesame was founded in 2016 by semiconductor industry veteran Jizhang Shan, who was previously with OmniVision for two decades. The company is positioned as an automotive-grade SoC provider from day one with core IPs like ISP and NPU developed in-house.</p><p>It launched the first chip product for intelligent driving named Huashan A500 in 2019 followed by <a href="https://www.blacksesame.com/en/huashan-a1000/">A1000</a> in 2020. Designed for driving and parking applications, it takes 3 years for Huashan A1000 to onboard production vehicles and another 2 years to reach volume adoption from OEMs like Dongfeng and Geely. The company is still shipping A1000 chips now. Some own it to the advanced architecture, and some believe Black Sesame had to catch up on toolchain offerings to support production development.</p><p>The A1000 product family has another two variants being <a href="https://www.blacksesame.com/en/huashan-a1000pro/">A1000Pro</a> and <a href="https://www.blacksesame.com/en/huashan-a1000l/">A1000L</a>, which are less popular. The A1000L has been deployed in two of FAW Hongqi models in 2024.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/415/1*oirItVx9ytawZyrc9yP1tA.png" /><figcaption>Huashan A1000 series</figcaption></figure><p>Speaking of vision on product design, Black Sesame launched Wudang C1200 in 2023 as the global first SoC for cross domain applications that can integrate ADAS, infotainment and vehicle control functions. In 2024, it released two variants of the C1200 family being <a href="https://www.blacksesame.com/en/wudang-c1236/">C1236</a> and <a href="https://www.blacksesame.com/en/wudang-c1296/">C1296</a>. The Wudang C1296 will be carried by the 2026 edition of Dongfeng Yipai 007 model enabling intelligent driving and smart cabin solutions.</p><p>In 2025, Black Sesame launched Huashan <a href="https://www.blacksesame.com/en/huashan-a2000/">A2000</a> series with optimized architecture to run mainstream AI models and multiple safety mechanisms for L2++ to L4 intelligent driving systems. The chip was reported to have passed relevant reviews by the US Department of Commerce and Department of Defense in January this year with permit for global sales and deployment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FMM3aMhMqm45w7mx1jtH-A.png" /><figcaption>Huashan A2000 series</figcaption></figure><p>The company has also established a new business line to provide computing platform solutions for robotics, the center of which is its proprietary NPU core. The robotics business has generated some revenue last year according to its annual report.</p><h3><strong>Axera</strong></h3><p>Axera was founded in 2019 by semiconductor industry veteran Xiaoxin Qiu, whose extensive research and development experience includes 15 years at Broadcom. The company is dedicated to design AI SoCs for edge scenarios with NPU and ISP cores developed in-house. It launched early chip products for vision-based terminal devices, the business line of which has made to top 5 globally.</p><p>The team started development for automotive-grade SoCs in 2021. In June 2023, Axera’s automotive chip named <a href="https://www.axera-tech.com/en/product/2824.html">M55H</a> first hit the road on Leapmotor vehicle models. As a cost-effective option for smart front cameras, M55H shipment ramped up in the following years with adoptions by GAC Aion, Geely Galaxy, Ford, etc.</p><p><a href="https://www.axera-tech.com/en/product/2826.html">M57</a>, an upgraded version of M55H, was launched in 2025 that can power smart front cameras or 5-camera based driving and parking solutions. Besides overall improved performance, it is embedded with safety MCU that can drive system cost down. It is also aligned with global regulatory and safety benchmarks for international deployment. New vehicle models using M57 will be released later this year according to the company.</p><p><a href="https://www.axera-tech.com/en/product/2825.html">M76H</a> was developed alongside M55H targeting L2+ ADAS systems to enable highway NOA, APA and possibly city ICA features with no OEM adoptions announced yet. The company will launch a 700+ TOPS M97 later this year aiming for L2++ systems with city NOA capabilities, forming a comprehensive chip portfolio.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*k-8g639EgCGwjof1MvuUAA.png" /><figcaption>Axera M57 and M95/97</figcaption></figure><p>Overall, Axera has demonstrated an unusual fast pace from product development to commercial deployment as a latecomer to the competitive ADAS SoC market. The founder Qiu owns it to its accurate product proposition addressing auto OEMs’ pain points and the tier2 commitment. It is also building chips for edge AI inference, another fast-growing direction of the company.</p><h3><strong>Huixi (Rhino Auto)</strong></h3><p>Huixi, also known as Rhino Auto, was established in 2022 to develop high performance autonomous driving chips by three Tsinghua alumni with two from the semiconductor industry and one’s background in ADAS.</p><p>It launched a 7nm chip R1 in 2024 with total compute power equaling two of Nvidia Orin-X. The company teamed up with HiRain, a local tier1 supplier, on a project to deliver full-stack city NOA solutions, in which HiRain was responsible for basic software and domain controller hardware and Huixi focused on chip design and application software. R1 faced some delay during the BIS review resulting in the OEM pulling back the plan to adopt Huixi chips, although R1 ended up passing the review.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/692/1*Z3K8XzQ_fMy3nb50dpN-mg.png" /><figcaption>Huixi’s R1</figcaption></figure><p>In 2025, the company has shifted to target humanoid robot applications with compute platform offerings based on the R1 chip. The new solutions have been commercially deployed and perhaps Huixi would re-visit its automotive strategy in the near future.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=413e54972d44" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[VLA and World Models Powered Autonomous Driving in China]]></title>
            <link>https://schen583.medium.com/vla-and-world-models-powered-autonomous-driving-in-china-96364c6ac2c2?source=rss-e5271e521780------2</link>
            <guid isPermaLink="false">https://medium.com/p/96364c6ac2c2</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[china]]></category>
            <category><![CDATA[reinforcement-learning]]></category>
            <category><![CDATA[autonomous-driving]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <dc:creator><![CDATA[Shuai Chen]]></dc:creator>
            <pubDate>Fri, 27 Mar 2026 09:10:24 GMT</pubDate>
            <atom:updated>2026-03-27T09:10:24.599Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*QIGTNgbznZL3mRAnUVR-kg.jpeg" /></figure><p>There are refreshed buzzwords in the China autonomous driving community every year. I wrote about <a href="https://medium.com/@schen583/chinas-4-stage-approach-to-end-to-end-autonomous-driving-b018b3c706ec">“end-to-end”</a> for the year of 2024 and since 2025, the new words are no doubt VLA (Vision-Language-Action) and world models.</p><p>While an end-to-end system could enhance the driving experience to a certain level, challenges remain such as fragile generalization in rapidly evolving scenarios and “black box” behaviors. The recent progress of multimodal large language models has inspired a new framework for autonomous driving involving VLA and world models to enable the system with reasoning capabilities as well as prior knowledge of driving and the space.</p><p>In this post, we look at some of the OEMs in China on their approaches with VLA and/or world models to further tackle autonomous driving.</p><h3>Li Auto</h3><p>Li Auto was the first to rave about incorporating a VLA model in its autonomous driving system while using a world model for cloud simulation back in early 2025, as in exhibit 1.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NL7V58NpScKWeZml3iebrA.png" /><figcaption>Exhibit 1: Li Auto at Nvidia GTC 2025</figcaption></figure><p><strong>V-Spatial Intelligence</strong> focuses on tokenizing sensor inputs and vehicle state to form 3D understanding of the environment via self-supervised Gaussian pre-training.</p><p><strong>L-Linguistic Intelligence</strong> performs inference with an LLM, which is trained from scratch with driving knowledge, reasoning capabilities and 3D spatial understanding. It has leveraged the MoE architecture, sparse attention and condensed CoT technologies to ensure real-time processing on the vehicle.</p><p><strong>A-Action Policy</strong> converts action tokens to trajectories through modeling the interactions between the vehicle and all traffic participates. It can take human instructions into consideration like taking the fastest route. It has adopted an ODE-based sampler for a more efficient diffusion process.</p><p>On the cloud, <strong>reinforcement learning</strong> is performed in simulation via a high-fidelity world model to further optimize the VLA model.</p><p>At the recent Nvidia GTC 2026, Li Auto introduced the next generation architecture named MindVLA-o1 per exhibit 2 to address the limitations of mainstream VLA approach.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*h2XBUgM0xOh-rK7oXedgAA.png" /><figcaption>Exhibit 2: Li Auto at Nvidia GTC 2026</figcaption></figure><p>Compared to the previous version, MindVLA-o1 will adopt a native multimodal transformer architecture where videos, languages and actions are pre-trained in a unified manner. It has further improved the 3D encoder via self-supervised pre-training with feedforward representations by decomposing the dynamics and statics. Besides the explicit linguistic reasoning capabilities, it will add a latent world model to predict future tokens as visual reasoning for enhanced planning. Trajectories will be generated through an action expert by aggregating key information from previous steps in a highly efficient way by applying parallel decoding and discrete diffusion.</p><p>Together with large scale driving data (MindData) and multimodal world model (MindSim) for training and testing (RL Infra) in a closed loop (exhibit 3), Li Auto envisions MindVLA-o1 as a framework to deploy artificial general intelligence in the physical world, where autonomous driving is one of the first application scenarios.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*i4qNjByjQ6Tvegz-7lnB5g.png" /><figcaption>Exhibit 3: Li Auto at Nvidia GTC 2026</figcaption></figure><h3>Xpeng</h3><p>At the CVPR 2025 WAD several months ago, Xpeng shared its philosophy on the next generation autonomous driving software with large foundation models and detailed on the current practices.</p><p>The architecture of Xpeng’s foundation model is per exhibit 4 with the backbone being an LLM and an action decoder. It takes input from videos and text prompt like navigation and ego position. Xpeng is using cameras only, no LiDARs or radars. The LLM engine processes all the tokens with CoT reasoning for action decoder to output trajectories.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xAdSgXGZsiJYM22MYPeHBA.jpeg" /><figcaption>Exhibit 4: Xpeng at CVPR 2025 WAD</figcaption></figure><p>Xpeng has identified three stages to train the foundation model (exhibit 5). First, it takes image and text pair from real driving scenarios to train an existing VLM (Qwen from Alibaba) for a driving-specific VLM. Then it takes video and action data set to do action pre-training and supervised fine-tuning for a driving-specific VLA model. Finally, it adopts reinforcement learning with rewards to do post-training for safer and enhanced driving behavior.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3vn1HyYLA_OERdFBN2Sqpw.jpeg" /><figcaption>Exhibit 5: Xpeng at CVPR 2025 WAD</figcaption></figure><p>Meantime, the company has built a closed loop system to ensure continued learning from “inner loop” to “outer loop” (exhibit 6). Inner loop includes model’s pre-training, post-training and deployment to production vehicles via distillation, which are performed in the office. Outer loop relies on the model collecting new data from real world for the next round training, with hundreds and thousands of production vehicles running on public roads.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zlUGC-la6JR4PgCBxyMohw.jpeg" /><figcaption>Exhibit 6: Xpeng at CVPR 2025 WAD</figcaption></figure><p>Earlier this month, Xpeng dedicated a press conference to launch VLA2.0 without prior information on VLA 1.0. It reiterated the encoder of its foundation model being a native multimodality tokenizer for accurate 3D understanding. It has been working on improving the efficiency of its visual reasoning capabilities to tackle complex driving scenarios in real time. The outputs include actions as well as videos and voices that can be later integrated with the smart cabin systems. Overall, Xpeng believes VLA2.0 will be the foundation to L4 autonomous driving as the team continuously iterating the model with a data-centric approach.</p><h3>Xiaomi</h3><p>Xiaomi claimed its HAD Enhanced intelligent driving software released in November 2025 has incorporated world model-based reinforcement learning to build up cognition capabilities via consumer fleets. During Nvidia GTC 2026, the company shared for the first time on its practices building a cognitive end-to-end system. The high-level system architecture is illustrated in exhibit 7.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mfQcY5HdqJXznnBCKxaiyQ.png" /><figcaption>Exhibit 7: Xiaomi at Nvidia GTC 2026</figcaption></figure><p>To utilize inputs from 11 cameras as well as 1 LiDAR and navigation with limited compute power, Xiaomi has identified three tiers of camera regions each running at different frame rates. It is also dynamically adjusting the frame rate of tier 1 sensors based on specific driving scenarios, as in exhibit 8.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*txt5dDrqk18it2w2gb32Qw.png" /><figcaption>Exhibit 8: Xiaomi at Nvidia GTC 2026</figcaption></figure><p>The core of its architecture is a student ViT backbone and latent features to do causal reasoning of the environment at low latency for vehicle deployment (exhibit 9). The lightweight ViT model is distilled from multiple VFM teachers on the cloud that have prior understandings of such as dense features, concept boundaries and 4D occupancy flow. Latent features are distilled from a pre-trained world model to provide generative world knowledge.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SopTiHQRS-gDnKw2IkBBBg.png" /><figcaption>Exhibit 9: Xiaomi at Nvidia GTC 2026</figcaption></figure><p>Then the 3D representations go through asynchronous fusion on BEV and a token re-sampler to form persistent cognitive tokens. Hierarchical generative planner eventually outputs the optimal trajectories using diffusion or autoregression. Reinforcement learning is conducted via the Xiaomi Genesis World Model on the cloud with human intervention data from consumer fleets being one of the key sources of reward mechanism.</p><p>It seemed that Xiaomi highlighted the efforts on vehicle deployment at efficient computational cost, probably for which reason it didn’t involve the language modality. However, the company did mention about research coverage on all technical paths including VLA at other occasions.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=96364c6ac2c2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ADAS and IVI Integrated Single Chip Solutions Finally in Production]]></title>
            <link>https://schen583.medium.com/adas-and-ivi-integrated-single-chip-solutions-finally-in-production-60aa3bdaebb7?source=rss-e5271e521780------2</link>
            <guid isPermaLink="false">https://medium.com/p/60aa3bdaebb7</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[china]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[electric-vehicles]]></category>
            <category><![CDATA[qualcomm]]></category>
            <dc:creator><![CDATA[Shuai Chen]]></dc:creator>
            <pubDate>Mon, 09 Feb 2026 09:10:02 GMT</pubDate>
            <atom:updated>2026-02-09T09:11:37.244Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*LQhnSOPtpL7z3OigP3YqyA.jpeg" /></figure><p>The ADAS and IVI integrated solutions have been discussed for a few years for the lower cost, simplified architecture and enhanced driving experience. It is also a variant along with the shift from distributed to centralized vehicle E/E architecture. Now it has gone from a concept on the roadmap to open public roads. There were two new models launched in the past Q4 of 2025 with ADAS and IVI systems consolidated on one chip, probably the first globally.</p><h4>BAIC Arcfox Alpha T5</h4><p>Launched in October 2025, the refreshed T5 is a mid-sized SUV offering 7 trim levels including extended range EVs and pure EVs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/415/1*9Er4T6-DZlX09tuEW2JeRA.png" /></figure><p>The trims priced above 130,000 yuan (about $19,000) are equipped with the ADAS and IVI integrated solutions built on a single Qualcomm SA8775P. It features Navigation on Autopilot (NOA) on highways and city roads as well as advanced parking assist leveraging 7 cameras, 3 radars and 12 ultrasonic sensors. Zhuoyu Technology/<a href="https://medium.com/@schen583/china-intelligent-driving-tier-1-dji-automotive-e1a6d44c9c03">DJI Automotive</a> supplied the ADAS stack from basic software, middleware to application software.</p><p>Smart cabin experiences include interconnected central control and instrument cluster, HUD multi-screen display, integrated AVM surround view and voice assistant supported by multiple onboard AI models, some of which are usually seen on mid to high-end vehicles.</p><p>Autolink supplied the cross-domain controller hardware and IVI system software. Founded in 2014, Autolink has become one of the leading intelligent cockpit solution providers on the China market with product offerings span from entry-level to high-end. The ADAS and IVI integrated single chip controller is among its latest innovations.</p><h4>Dongfeng Nissan N6</h4><p>N6 is a brand-new model launched by the Dongfeng and Nissan joint venture in December 2025. It is a plug-in hybrid mid-sized sedan with five trim levels. The high trims, 180 Max+ and 170 Max+, adopt the ADAS and IVI integrated solutions built on a single Qualcomm SA8775P, which lowered the threshold of high performance single-chip solutions to the price segment of 120,000 yuan (about $17,000).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/415/1*tMhIDTc48Phf0C8NLcIClw.png" /></figure><p>The ADAS software was from Momenta. It is capable of highway NOA, city memory NOA and advanced parking assist via 7 cameras, 3 radars and 12 ultrasonic sensors. City memory NOA means the system builds up capabilities on frequently driven routes after certain amount of time.</p><p>The cross-domain controller hardware was provided by Hangsheng. Also known as HASE, the company is involved in the development and production of automotive electronics since 1993. With the demand for vehicle intelligence in recent years, it has transformed into a major supplier of IVI components and system solutions.</p><h4>ADAS and IVI integration: one-box, one-board and one-chip</h4><p>China has been raving for ADAS and IVI integrated solutions driven mainly by heightened cost pressures within the region. An early practice is to integrate the driving PCB board (with driving SoC) and the cockpit PCB board (with cockpit SoC) in one-box, two boards communicating via Ethernet. A further step is to integrate driving SoC and cockpit SoC on one PCB board, namely one-board. Eventually, it can be a single SoC to handle both driving and cockpit functionalities, namely one-chip.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/807/1*PoZo9IUWGao4AfM0fzS6Og.jpeg" /></figure><p>OEMs would favor the ultimate one-chip approach for reduced BOM cost, 20–30% down on hardware in the cases of Alpha T5 and N6. It could optimize power consumption, especially important to battery EVs, and foster streamlined communication between two systems, all leading to more powerful and responsive end-user experience.</p><p>However, it is a bit challenging during the R&amp;D stage, for example, to isolate safety-critical driving functions from user-facing digital experiences and to innovate on software architecture to allow dynamic compute resource allocation.</p><p>For now, single-chip solution is targeting mainstream vehicles priced between 100,000 and 150,000 yuan ($14,000–22,000). Qualcomm’s current 8775 and next generation 8797/8397, purposely designed for merging infotainment and ADAS, offer adequate level of compute at a fraction of the cost compared to two-chip setup. Mid to high-end models would still adopt multi-chip setups with redundant compute to feature AI chatbots, larger and higher-resolution displays such as “pillar-to-pillar” panoramic screens and 3D/AR displays, VLM/VLA models in the driving system, etc.</p><h4>Single-chip solutions to disrupt the mainstream market?</h4><p>Intelligent cockpit has achieved nearly 80% penetration rate on mainstream vehicles (priced 100,000–150,000 yuan), and a popular configuration is Qualcomm 8155. It usually pairs with a smart front camera enabling regulatory L2 functionalities built on Horizon J2/3 or Mobileye chips. Optioning for sing-chip solutions would bring upgraded intelligence, highway NOA and potentially city NOA, to the vehicle models without much additional cost.</p><p>The one-chip approach also pushes for a new collaboration model among OEMs, IVI suppliers and ADAS suppliers. In the case of Arcfox Alpha T5 and Nissan N6, Autolink and Hangsheng that specialize in IVI domain controllers were the tier1 performing overall system integrations whereas Zhuoyu and Momenta became tier2. It would be a necessary trio to launch single-chip solutions unless one supplier has expertise in both IVI and ADAS.</p><p>Desay SV, with capabilities in both domains, is reported to be building single-chip solutions on Qualcomm platform for Chery and JAC Ford models. Once the projects are market-proven, it would be Desay’s edge to win further shares on the mainstream market. Bosch obviously can adopt the same strategy. It claims to be working on single chip solutions but not much details have been revealed by far.</p><p>2025 marked the first attempt with single-chip solutions based on Qualcomm 8775. There will be solutions based on the more powerful 8397 and 8797, as well as on other chips like Black Sesame C1296. As for whether it would disrupt the market of smart front cameras and positions of IVI and ADAS solution providers, we have the next two to three years to find out.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=60aa3bdaebb7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[China Automakers and Tier1s’ In-House ADAS Chips to Challenge Nvidia Thor/Orin]]></title>
            <link>https://schen583.medium.com/china-automakers-and-tier1s-in-house-adas-chips-to-challenge-nvidia-thor-orin-4debcc9bd3c3?source=rss-e5271e521780------2</link>
            <guid isPermaLink="false">https://medium.com/p/4debcc9bd3c3</guid>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[automotive]]></category>
            <category><![CDATA[china]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[nvidia]]></category>
            <dc:creator><![CDATA[Shuai Chen]]></dc:creator>
            <pubDate>Mon, 19 Jan 2026 09:40:26 GMT</pubDate>
            <atom:updated>2026-01-19T09:40:26.122Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/864/1*v-P-PTyYO61wsk_mGzR8aA.png" /><figcaption>Source: WeChat Account — More Then Semi</figcaption></figure><p>A number of Chinese OEMs and Tier1s have been quietly designing their own ADAS chips for years, as a strategic move to achieve self-sufficiency, better hardware utilization than off-the-shelf solutions and eventually optimized BOM cost, though risky at the same time given the high upfront investment. 2025 has witnessed two of the in-house chips onboarding production models and some close to hitting the road, among many other momentums in the autonomous driving sector. It is time for an overview.</p><h3>OEMs</h3><h4><strong>Nio’s Shenji NX9031</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Mt9rSCK1ASb_xMS03n5_eQ.png" /><figcaption>NX9031 marketing slide</figcaption></figure><p>Nio first introduced its self-developed chip for high-level intelligent driving system codenamed Shenji NX9031 in late 2023 while the efforts dated back to 2021. With the delivery of ET9 in March 2025, the first model powered by Shenji chips, it has officially reached production stage.</p><p>NX9031 uses a 5nm process technology with over 50 billion transistors and delivers 1,000+ TOPS compute power. With a 32-core design, it features the ARM big.LITTLE configuration with combined Cortex-A78 and A55 processor units to balance power efficiency and performance.</p><p>Nio CEO once referred to the R&amp;D cost on NX9031 as building 1,000 battery-swap stations. The project involves over 600 engineers covering front-end design, back-end design, verification and testing, the scale of a chip design house. However, it brings a BOM optimization of about 10,000 yuan (about $1,400) per vehicle, according to the company.</p><p>Meanwhile, Nio is suspected to be seeking potential revenue stream by outsourcing the technology. In June 2025, it took a first step by spinning-off the chip division into Anhui Shenji Techoloogy Co., Ltd. In November, the chip subsidiary formed a joint venture with OmniVision and Axera Semiconductor, with Axera being the largest shareholder. Axera is an emerging rival to Horizon Robotics in China on ADAS chips.</p><h4><strong>Xpeng’s Turing</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/907/1*Ah3hnuOL6u4EoLWFyQuFGA.png" /><figcaption>Xpeng CEO introducing Turing</figcaption></figure><p>Xpeng once disclosed that it took them 5 years to launch the Turing chip on G7 Ultra version in July 2025. Turing also adopts the 5nm process technology with approximately 750 TOPS compute power. Despite limited official information, some tech geeks made presumptions on Turing’s “all big core” design for high efficiency and low cost, which is quite different from Nio. It is a 40-core configuration with ARM Cortex-A78 CPUs, Mali-G78 GPUs and Cortex-R52 cores in lockstep. It features an integrated dual-NPU, probably one of which for high load tasks and the other for low compute tasks. There are two independent ISPs designed for the differentiated camera configurations of vehicle models.</p><p>A Turing trio currently powers the Ultra version of newest Xpeng models including 2026 edition of G6, G9, G7 and P7+, where ADAS system adopts two Turing chips and infotainment system pairs one Turing with Qualcomm SA8295P or MediaTek 8676. As Turing is tailored particularly for running large AI models, Xpeng envisions the same chips and AI models to power its eVOTL/flying cars, humanoid robots, and 2026 to-be-launched robotaxi.</p><p>As is known, Xpeng has fostered a strategic partnership with Volkswagen since 2023 primarily on joint development of E/E architecture for VW-branded models in China and generated substantial income by providing technical services. The scope of the collaboration has been extended recently for Xpeng to license ADAS technologies including the Turing chips and its VLA models.</p><h4><strong>Geely and SiEngine’s AD1000</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_zD58R-YuwgMY_t_sbUJEA.png" /><figcaption>SiEngine smart cockpit solutions</figcaption></figure><p>SiEngine was established in 2018 with investment from the Geely-backed EcarX and Arm China to develop automotive AI chips. The company first launched SE1000 in 2023, a 7nm purpose-built chip for infotainment systems, comparable to Qualcomm 8155. The chip has now been widely adopted by Geely sub-barnds as well as certain models from FAW Hongqi, Changan Qiyuan and VW.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YbJZ7SZjbR5u-nyjuDU6aw.jpeg" /><figcaption>SiEngine intelligent driving solutions</figcaption></figure><p>Last month, SiEngine announced that its 7nm ADAS chip dubbed AD1000 has reached mass production with automakers and ecosystem partners working on deployment. AD1000 brings up to 512 TOPS in NPU capacity and can reach a maximum of 2,048 TOPS in a cascaded manner, providing sufficient compute power for L2 to L4 systems. It features a multi-core heterogeneous architecture optimized for vision processing and mainstream AI frameworks.</p><p>SiEngine has laid out a product roadmap with the next generation infotainment and ADAS chips, all geared up to battle with imported solutions. At the moment, EcarX is still the largest shareholder while a list of OEMs like FAW group, Tier1s and semiconductor value chain players all hold stakes in the company, revealing SiEngine’s ambition to build up a neutral and open platform.</p><h4><strong>Leapmotor’s Lingxin 01</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*sawmuPhMdfgHUBoa--DzyQ.png" /><figcaption>Lingxin 01 marketing slide</figcaption></figure><p>Leapmotor is in fact the first OEM devoted to in-house ADAS chips back in 2018. The team designed Lingxin 01 with “Xuantie C860” CPU processor from T-Head Semiconductor (Alibaba subsidiary) as an alternative to the much-demanded Mobileye EyeQ4 at the time. The 28nm Lingxin 01 was capable of 4.2 TOPS compute power on its own, two of which powered the driving and parking features on Leapmotor’s C11 model first delivered in 2021. It was reported that vehicles equipped with the self-supplied chips had reached 100,000 in volume by December 2024. However, the company has shifted strategies to embrace vendors like Qualcomm for ADAS systems in recent years.</p><h4><strong>Other OEMs</strong></h4><p>Li Auto, Xiaomi, BYD and Chery are said to be developing ADAS chips as well. <strong>Li Auto</strong> is working on ultra-high computing solutions like Nio and Xpeng to better support its VLA model, and the chip is close to production stage. <strong>BYD</strong> is probably aiming for Horizon J6E alternatives to power its mainstream models priced between 100,000–200,000 yuan ($14,000–28,000). No details revealed on Xiaomi and Chery.</p><h3><strong>Tier1s</strong></h3><h4><strong>Huawei</strong></h4><p>Huawei has evolved to the 4th generation with its ADS solutions, all of which are powered by the Ascend 610 chips from HiSilicon, a Huawei subsidiary company. Ascend 610 is purpose-built for intelligent driving applications. It adopts a 7nm process with 200 TOPS compute power each. Based on Huawei’s proprietary Da Vinci architecture, the chip features high performance, high power efficiency, expandability and robustness when it comes to AI computing.</p><h4><strong>Momenta</strong></h4><p>Momenta is a major ADAS Tier1 in China, known for its production-proven “city NOA” capabilities for high-end vehicle models. The company was reported to be developing in-house chips since 2023. Although not officially confirmed, the chip business will be conducted by an independent entity named Xinxin Hangtu Technology Co., Ltd. Sources indicate its first product will be in the same segment with Nvidia Orin-X or Qualcomm 8650 delivering a compute power of 272 TOPS. It is designed with similar interface as 8650, making a potential swap feasible without major changes to the PCB boards.</p><p>Momenta CEO counted 300,000 vehicles carrying its city NOA solutions as of May 2025, with the units growing in a rapid speed. The company currently offers a solution built on Orin-X mainly for local brands, and another solution on Qualcomm platform to supply international brands as well as for combustion engine cars. Once the in-house chips are in place, Momenta would be able to offer full-stack software and hardware optimized solutions at theoretically competitive price. As OEMs seek to democratize city NOA features onto mainstream cars this year with cost being primary concerns, Momenta could be well positioned to tackle a new market.</p><h4>Baidu</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*b4Yty23P-7DXmgqJ_nEQ2g.png" /><figcaption>Kunlunxin Gen 2 marketing slide</figcaption></figure><p>Baidu, strictly speaking, is neither an auto OEM nor a Tier 1. It is making the headlines lately as its chip subsidiary Kunlunxin has filed a Hong Kong IPO with anticipated fast growing business from 3rd party customers as well as Baidu internal demand. The 2nd generation of Kunlun AI chip has been powering Apollo robotaxis since 2022, currently deployed in multiple cities of China. It is a 7nm chip with 256 TOPS compute power built on the proprietary XPU-R architecture.</p><p>According to Kunlunxin’s product roadmap, Baidu is mainly eyeing the chip market for data centers and cloud computing, less on automotive grade solutions.</p><p>Last but not least, local independent semiconductor companies are a strong force on the China ADAS chip market, driving innovations at affordable cost. I will have another post on it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4debcc9bd3c3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Huawei Intelligent Driving Solutions Overview — A $16 Billion Valued Automotive Tier 1]]></title>
            <link>https://schen583.medium.com/huawei-intelligent-driving-solutions-overview-a-16-billion-valued-automotive-tier-1-f88f4bdfdff9?source=rss-e5271e521780------2</link>
            <guid isPermaLink="false">https://medium.com/p/f88f4bdfdff9</guid>
            <category><![CDATA[autonomous-driving]]></category>
            <category><![CDATA[china]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[huawei]]></category>
            <category><![CDATA[autopilot]]></category>
            <dc:creator><![CDATA[Shuai Chen]]></dc:creator>
            <pubDate>Wed, 04 Dec 2024 05:28:17 GMT</pubDate>
            <atom:updated>2024-12-04T05:28:17.769Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Huawei Intelligent Driving Solutions Overview — A $16 Billion Valued Automotive Tier 1</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*0J3Xb7KdZGwBr5vdigicmQ.jpeg" /></figure><p><em>This is the 2nd of a series to introduce new generation Tier 1s for intelligent driving solutions in China. Read the 1st one </em><a href="https://medium.com/@schen583/china-intelligent-driving-tier-1-dji-automotive-e1a6d44c9c03"><em>here</em></a><em>. My recent posts focused on </em><a href="https://medium.com/@schen583/china-intelligent-driving-tier1s-leading-the-end-to-end-paradigm-shift-e545fa9380f1"><em>end-to-end autonomous driving</em></a><em> and </em><a href="https://medium.com/@schen583/the-stakeholders-shifting-china-automotive-lidar-market-b6774607a3bd"><em>automotive LiDAR</em></a><em>, in all of which Huawei had to be mentioned. The company has obviously become a heavyweight of the growing EV/AV industry that requires a dedicated post.</em></p><p>In retrospect, Huawei has achieved a milestone on the mission “to do everything short of building a vehicle”, quoting my original words in a <a href="https://medium.com/@schen583/huaweis-auto-ambitions-aa481e9f9222">post</a> from 2019. While some of the tech companies in the field struggled to survive, Huawei’s automotive business unit (BU) recorded a net profit of CNY2.231 billion in H1 of 2024 from a net loss of CNY5.597 billion in 2023. Now the automotive BU has been spun off into an independent subsidiary company, Shenzhen Yinwang Intelligent Technology Ltd., with a CNY115 billion (USD16 billion) valuation.</p><p>Huawei’s Yinwang currently has five business lines that are: Qiankun intelligent driving, Harmony cockpit, Qiankun auto lighting, Qiankun vehicle control, and Qiankun cloud service. We dive into Qiankun intelligent driving on its solutions and iterations as well as the different collaboration models with auto OEMs.</p><h3><strong>ADS solutions</strong></h3><p>To straighten out some terms, Navigation Cruise Assist (NCA) is Huawei’s term for the commonly used Navigation on Autopilot (NOA). Similar to Tesla FSD, it is an advanced intelligent driving function allowing a car to navigate itself on the road with actions like stopping, steering, changing lanes, etc. It has been implemented for highways and city expressways on a scale and is rolling out to complex city roads. So, city NCA is a level up from highway NCA implicitly including the latter.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*8Yz1QeLUii6JF4S4z2DUXw.png" /></figure><p>Huawei’s intelligent driving solution was first adopted by BAIC Arcfox Alpha-S Hi version in 2021, followed by Avatr 11 in 2022. Now referred as ADS 1.0, the system hosted a prestigious set of sensors on two of Ascend 610 chips with 400 TOPS compute power in total. With transformer-based BEV architecture and the input from HD map, it allowed vehicles to navigate in city traffic scenarios. However, city NCA could only be activated in a few cities of China due to the limited availability of HD map. The system cost was estimated to be about CNY100,000 while the vehicle models were priced around CNY400,000 at the time.</p><p>The 2nd generation was initially launched with AITO M5 in 2023, followed by a few other models. It implemented GOD (General Obstacle Detection) and RCR (Road Cognition &amp; Reasoning) on the software level to replace the functions of HD map and also downsized the hardware suite by removing 2 cameras, 3 mmWave radars, 2 LiDARs and 1 Ascend 610 chip. By the end of 2023, it was announced that ADS 2.0 has realized city NCA all over China. Meanwhile, the system cost was optimized to about CNY30,000.</p><p>The <strong>ADS 3.0</strong> was launched earlier this year on Stelato S9 and then made available to precedent models for upgrade. It incorporated neutral networks in the perception and planning modules for smooth human-like driving experience. As a step up from the previous version, city NCA can be activated from parking space and lead all the way to the next parking space. The system configuration remained the same with ADS 2.0 except the LiDAR and mmWave radars were updated to the newest model.</p><p>For end consumers to access the software pack for city NCA and AVP, they have the options to do an upfront payment of CNY36,000, monthly subscription of CNY720 or yearly subscription of CNY7,200, discount applicable during certain periods. Highway NCA and APA are standard features without additional cost.</p><p>To meet the growing demand of intelligent driving on affordable vehicles, Huawei recently introduced a basic-level solution called <strong>ADS SE</strong>. It skipped the LiDAR sensor and a rear view camera from the ADS 3.0 configuration, supporting highway NCA and APA by default. To access advanced features like city LCC Plus and AVP, it requires an upfront payment of CNY5,000 or subscriptions of CNY100 per month or CNY1,000 per year. It has been adopted by two models from Changan Deepal as well as the Pro edition of Aito M7, Luxeed S7 and R7, which didn’t come with ADS 3.0.</p><h3><strong>Vertical integrated industry chains</strong></h3><p>Looking into the supply chain of ADS solutions, Huawei owns some of the core technologies, often referred as a vertical integration of chip-hardware-software-cloud.</p><p>The MDC controller platform is built on Ascend AI chipset from HiSilicon, a subsidiary company of Huawei, with tailored operating system, middleware and tool chain.</p><p>In terms of sensors, Huawei’s LiDAR product has evolved from D2 of 92-channel launched in 2020 to D3 of 192-channel launched this year. The next generation D5 of 512-channel will be unveiled in 2025. Although not widely discussed, Huawei has jumped to the top supplier chart for front radar (including 4D) and corner radar, rapidly gaining market share this year.</p><p>As the industry is shifting towards the end-to-end architecture, data and computing power are crucial for any model-based intelligent driving systems to scale. In October, Huawei recorded a total computing power of 7.5 EFLOPS, allowing a model iteration every five days trained with 35 million km data on a daily basis, all of which are made possible with Huawei Cloud.</p><h3><strong>Three collaboration models with auto OEMs</strong></h3><p>Over the years, Huawei has identified three models to work with auto OEMs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/867/1*wve5KZVTjAfWRAf5J0SQLg.png" /></figure><p>The first model is called <strong>Harmony Intelligent Mobility Alliance (HIMA)</strong>, in which Huawei acts as a Tier 0.5 supplier. It is involved in the entire life cycle of a vehicle model by empowering the auto OEM from product design and development, quality and supply chain management, branding, online &amp; offline sales to after sales service. Huawei has launched four brands with four OEMs under this model that are AITO with Seres, Luxeed with Chery, Stelato with BAIC and Maextro with JAC.</p><p>HIMA has achieved significant commercial success with over 500,000 vehicles delivered in 32 months, although this year contributed a big part of the sales volume. The average price of all HIMA models was rounded to CNY 399,000 based on the data from January to August 2024. The brands under HIMA are perceived by consumers as high-end and high-level intelligent associated with the widely recognized Huawei image.</p><p>The second model is called <strong>Huawei Inside (HI)</strong>, where Huawei acts as a Tier 1 providing full stack hardware and software integrated solutions. For example, it has launched the Alpha-S vehicle model with BAIC Arcfox, three models with Changan Avatr, two models with Changan Deepal and a model with Dongfeng Voyah. Another model with M-Hero is expected to hit the market in Q1 of 2025.</p><p>The third model is for Huawei to supply standard parts or modules as Tier 2. It has over 30 products including all types of sensors, T-box, AR-HUD, powertrain, cloud service, etc. A longer list of vehicle models is under this category all containing auto parts from Huawei.</p><p>Meantime, Huawei is probably eyeing a bigger OEM-circle not restricted by the above-mentioned three models. Bao8 from BYD sub-brand Fangchengbao has been released with ADS 3.0 intelligent driving solutions, which is neither under HIMA nor HI strictly speaking. Also, it was reported that Trumpchi, a brand under GAC would launch three new vehicles in 2025 integrated with ADS 3.0 and Harmony cockpit solutions.</p><h3><strong>Yinwang to disrupt the automotive ecosystem?</strong></h3><p>Going back to Yinwang’s financials, the net profit of 2024 is expected to reach CNY5 billion with H1 net profit as CNY2.231 billion. The CNY115 billion valuation makes its PE multiple 23, which is on par with other Tier 1s such as Desay SV at 25 and Adayo at 20 (according to China Fortune Securities’ estimation of 2024). However, Yinwang is not merely a traditional Tier 1.</p><p>The automotive industry is facing severe challenges with intensive competition among new EV brands, high R&amp;D costs and price wars spreading to upper stream players. Huawei envisions an open platform enrolled by both OEMs and suppliers with a dedication to intelligent system solutions while covering design, development, production, sales and services, eventually beneficial to all parties.</p><p>The HIMA practice known as Tier 0.5 has proven to be successful and Yinwang is taking a further step to strengthen the alliance via capital. Before the entity was established, Huawei extended invitations to its existing and potential OEM partners, with Avatr and Seres being the early birds each acquired a 10% stake. More OEMs are expected to get on board. With Yinwang’s new role emerging, would it ultimately disrupt the automotive ecosystem? We have the next few years to find out.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f88f4bdfdff9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The “Stakeholders” Shifting China Automotive LiDAR Market]]></title>
            <link>https://schen583.medium.com/the-stakeholders-shifting-china-automotive-lidar-market-b6774607a3bd?source=rss-e5271e521780------2</link>
            <guid isPermaLink="false">https://medium.com/p/b6774607a3bd</guid>
            <category><![CDATA[china]]></category>
            <category><![CDATA[electric-vehicles]]></category>
            <category><![CDATA[lidar]]></category>
            <category><![CDATA[autonomous-cars]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <dc:creator><![CDATA[Shuai Chen]]></dc:creator>
            <pubDate>Wed, 30 Oct 2024 16:41:38 GMT</pubDate>
            <atom:updated>2024-10-30T16:41:38.243Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/660/1*zdTIN1fw0-2RsYmFsH908A.jpeg" /></figure><p>There was a time when automakers in China touted their EV models by stressing the number of LiDAR sensors and even the performance parameters of that LiDAR product. Regardless of whether such campaigns made sense to end consumers, we don’t hear this any more. Actually, we come across statistics such as a breakthrough in Q3 2024 with 423,400 LiDAR sensors delivered (for pre-installed passenger cars excluding exports).</p><p>In order to understand how the market has shifted over the past few years, we review the major “stakeholders” that will ultimately define the future of automotive LiDAR in China.</p><p>To straighten out some terms, Navigation on Autopilot (NOA) is an advanced intelligent driving function allowing a car to navigate itself on the road with actions like stopping, steering, changing lanes, etc. It is a similar concept to Tesla FSD but the naming was slightly twisted by different companies. For example,</p><p>· Huawei calls it NCA (Navigation Cruise Assist),</p><p>· Nio calls it NOP (Navigate on Pilot),</p><p>· Xpeng calls it NGP (Navigation Guided Pilot).</p><p>We refer it all as NOA in this post. Generally speaking, NOA has been implemented for highways and city expressways on a scale and is rolling out to complex city roads. So, city NOA is a level up from highway NOA implicitly including the latter.</p><h3><strong>BYD</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jHAfoCXX6uwWqLg0ELxpAw.png" /></figure><p>BYD was recently ranked the 7th world’s largest automaker by Q2 sales volume, surpassing Honda and Nissan for the first time. It seized the global BEV sales crown in Q4 of last year. The company has obviously skyrocketed within a few years time by betting on electrification. When US and China rivals were raving about the future of autonomous driving, the CEO called it a false proposition. Now it is burning over 100million yuan on a daily basis, according to the company’s first half of 2024 report, building up its core competence in intelligence. BYD has joined the race, just in time.</p><p>Yangwang is positioned as a luxury brand and Denza as mid to high-end brand in the BYD group, with each having launched a model in 2023 equipped with LiDAR sensors. This year, LiDAR seems to have become a standard sensor on Yangwang and Denza cars with the launch of Z9GT, upcoming U7 and all-new D9. Meanwhile, mass-market models from the Ocean and Dynasty series began to feature LiDARs. The 2025 Seal from the Ocean series launched a LiDAR edition in August priced 216,800+ yuan, followed by the 2025 Han from the Dynasty series with the LiDAR edition sold for 225,800+ yuan.</p><p>Basically, the LiDAR models come with DiPilot 300 to enable city NOA and are priced above 200,000 yuan. While the mid to high-end models are transforming with intelligence, BYD revealed a strong desire to offer advanced driving features to models sold between 100,000 to 200,000 yuan in 2025 and even below 100,000 yuan in the near future.</p><p>To achieve the ambition, BYD is open to work with different suppliers besides the growing in-house R&amp;D team. An upcoming model (Bao8) from the sub-brand Fangchengbao was confirmed to adopt Huawei ADS 3.0 intelligent driving solution, sold in the price range of 400,000 to 500,000 yuan.</p><p>Additionally, BYD was reported to be developing LiDAR sensors since 2022 aiming for the low-channel product with a targeted unit cost of 900 yuan. Given the company’s main battlefield is in affordable mass-market models, it makes sense to demand such LiDARs in its future sensor suite.</p><h3>Huawei</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*i4Ho3q24HBBBfMFeZCUsGQ.png" /></figure><p>As an intelligent driving solution provider, Huawei has teamed up with eight OEMs so far and together delivered various vehicle models, some of which got into top selling charts.</p><p>Its 1st generation solution was launched on Arcfox S and Avatr 11 in 2021. That leveraged 13 cameras, 6 mmWave radars, 12 ultrasonic sensors and 3 LiDARs allowing cars to navigate-on-autopilot on city roads that had HD map coverage. In 2023, ADS 2.0 was introduced and first carried by AITO M5 that no longer relies on HD maps. Thanks to the implementation of RCR (road cognition &amp; reasoning) on the software level, city NOA works all over China on structured roads with clear boundaries. Meanwhile, the sensor suite was downsized to 11 cameras, 3 mmWave radars, 12 ultrasonic sensors and 1 LiDAR.</p><p>The ADS 3.0 solution was launched this year to further improve driving experience by incorporating neutral networks in the perception and planning modules. It was implemented by the newest models such as Stelato S9 and AITO M9 and shortly made available to precedent models for upgrade.</p><p>From the 1st generation to ADS 3.0, Huawei has stuck to the sensor fusion approach with LiDARs. The company’s senior executive Richard Yu made it clear that LiDAR is a crucial part of their system for safety reasons, as the limitations of cameras would not be solved by algorithm. It wasn’t just a verbal statement because Huawei has upgraded its LiDAR technology with the all-new 192-channel product from the 92-channel model (reference <a href="https://schen583.medium.com/huawei-lidar-preview-922dd2d72828">post</a>).</p><p>It is worth noting the vehicles equipped with ADS 3.0 are currently sold above 200,000 yuan, mostly above 250,000 yuan. Besides the premium cost on hardware, users need to pay to access city NOA and AVP features through monthly subscriptions or an upfront purchase (highway NOA is free by default). However, Huawei has its sights on the mass-market too. It introduced a basic-level intelligent driving solution called ADS SE with highway NOA and APA functions without using LiDARs. It was first carried by Deepal L07 pro version, a vehicle model sold for 166,900 yuan.</p><h3><strong>New-gen EV trio</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*J_I7sqIrnex4D6jFlPgxEg.png" /></figure><p>Nio, Li Auto and Xpeng are often referred as the EV trio. Despite their slightly different market propositions, the trio waw the early adopter of many innovative technologies including LiDAR as well as pioneering EV manufacturing as startups on a regulated market.</p><p>Li Auto designs its vehicles for longer family trips with at least five seats and extra interior space. It currently has five vehicle models, each offering Pro, Max or Ultra editions. The intelligent driving system on Max or Ultra enables city NOA functions, which was built by internal team on two Orin-X chips with a sensor suite including one LiDAR. The system on Pro only supports highway NOA and was built by a supplier on a single Horizon J5 chip with fewer amounts of sensors not including LiDARs.</p><p>Xpeng centers its brand image around the use of high tech to attract Millennials, Gen Z and tech geeks in particular. It was one of the first to mass-produce vehicles with two LiDAR sensors back in 2021. Xpeng used to provide a Max edition on all its models with city NOA solution, which was built on two Orin-X chips with two LiDAR sensors. However, a new strategy might have been implemented starting with the P5 2024 model cancelling the Max edition, which means P5 doesn’t use LiDARs at all nor offers city NOA. Recently, it was confirmed for the upcoming P7+ model to adopt pure vision solutions by removing LiDARs but still using radars. Furthermore, Xpeng plans to upgrade existing models to pure vision in 2025.</p><p>Nio has long envisioned itself as a premium brand and strives for high standards from vehicle configuration to customer service. It has iterated all the models onto the NT2.0 platform, where the intelligent driving system is built on 4 Orin-X chips and stacked with 11 cameras (7 8MP cameras and 5 3MP cameras), 5 radars, 12 ultrasonic sensors and a 1550nm-wavelengthed LiDAR. In other words, all Nio cars come with the necessary hardware for city NOA but users have to pay a monthly subscription fee to access the software pack. This is different from Li Auto and Xpeng, with whom city NOA is available on selected vehicle editions while there is no additional cost to access it.</p><h3><strong>Paving the way for city NOA</strong></h3><p>It is only a matter of time before all cars on the road to become smart and then smarter, just like our mobile phones. A smart car can drive itself on highways and expressways with minimum driver’s input, which is no longer a hurdle technically. The focus has moved onto sophisticated city traffic scenarios, where the current solutions all rely on at least one LiDAR sensor.</p><p>In fact, premium new energy vehicles today are adopting city NOA features rapidly since cost isn’t the biggest concern, whereas affordable models are usually the mainstream choice for consumers in China and globally. OEMs and solution suppliers are racing to democratize the same experience to vehicles priced below 200,000 yuan, a segment that possesses abundant demand and growing market share.</p><p>There are two camps working on transferring high-level intelligent driving to lower price points. One is to go with pure vision by ditching LiDAR. Right now, only Xpeng is in this camp until Tesla FSD makes its way into China. The other camp sticks with sensor fusion hoping to optimize cost overall. There doesn’t have to be a single winner. In the foreseeable future, the automotive LiDAR market will stay at least on an upward curve.</p><p>Last but not least, the discussions above are from the demand perspective. I will probably do another post from the supply side.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b6774607a3bd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[China Intelligent Driving Tier1s Leading the End-to-end Paradigm Shift]]></title>
            <link>https://schen583.medium.com/china-intelligent-driving-tier1s-leading-the-end-to-end-paradigm-shift-e545fa9380f1?source=rss-e5271e521780------2</link>
            <guid isPermaLink="false">https://medium.com/p/e545fa9380f1</guid>
            <category><![CDATA[electric-vehicles]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[china]]></category>
            <category><![CDATA[autonomous-driving]]></category>
            <dc:creator><![CDATA[Shuai Chen]]></dc:creator>
            <pubDate>Wed, 18 Sep 2024 09:22:47 GMT</pubDate>
            <atom:updated>2024-09-18T09:22:47.868Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nnnYiQVnDbd6Z0v8mlsKfg.png" /></figure><p>In this post, we look at a number of intelligent driving Tier1s in China that have revealed notable progress in developing end-to-end solutions.</p><h3><strong>Huawei</strong></h3><p>Huawei has become a leading supplier of smart car technologies in just five years since it revealed its <a href="https://medium.com/@schen583/huaweis-auto-ambitions-aa481e9f9222">auto ambitions</a> in 2019. Speaking of intelligent driving, it was the first to make city NOA accessible across the entire country.</p><p>First launched in 2023, Huawei ADS2.0 adopted a General Obstacle Detection (GOD) network in the perception layer and a Road Cognition &amp; Reasoning (RCR) network to replace HD map, together enabling city Navigation Cruise Assist (NCA), also known as NOA. In the latest ADS3.0, it has aggregated the perception task to a single GOD network with scenario comprehension capabilities (exhibit 1). It has also incorporated a single neural network in the Prediction, Decision and Planning (PDP) layer. In parallel to the PDP network, there is an instinct safety network with pre-defined rules to ensure safety bottom line (exhibit 2). ADS3.0 was claimed to bring human-like driving style with the new end-to-end architecture.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*dLT7jlIo1ZrvC6M-7cENYQ.png" /><figcaption>Exhibit 1: Perception module evolution, source: Huawei press conference</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*qFawXI4MTfehBb5x0mcnSQ.png" /><figcaption>Exhibit 2: ADS 3.0 architecture, source: Huawei press conference</figcaption></figure><p>According to a recent press conference, Huawei has a total computing power of 5 EFLOPS, grown from 3.5 EFLOPS in June. The system is being trained with 35 million km data on a daily basis thanks to the rapid growing cars on the road fitted with ADS. ADS3.0 was first released on Stelato S9 (Stelato is a new brand under BAIC), and would be updated onto vehicle models from AITO, a brand under Seres, Luxeed, a brand under Chery, Avatr, a brand under Changan, etc.</p><h3><strong>Deeproute.ai</strong></h3><p>Deeproute started out in 2019 to develop L4 autonomous driving technologies. The company was less raved about as opposed to its robotaxi rivals like Pony.ai and WeRide until September 2021 when Alibaba led its series B funding round. Deeproute did launch a few L4 solution packages but has over time come to embrace L2 technologies, where the demand is much higher.</p><p>Last year, it introduced DeepRoute-Driver 3.0 for mass production passenger cars, which included hardware and software stack to enable city NOA without HD map reliance. In April of this year, it launched DeepRoute IO, an end-to-end solution built on Nvidia DRIVE Orin with 7 cameras and 1 LiDAR. According to the company’s marketing slide (exhibit 3), Deeproute has reached the one-model end-to-end architecture compared to most players in the modular end-to-end stage (reference <a href="https://medium.com/@schen583/chinas-4-stage-approach-to-end-to-end-autonomous-driving-b018b3c706ec">post</a>). The founder and CEO, in an interview, owned it to the team’s forward vision on autonomous driving technologies, due to which reason the development efforts have always been ahead of buzzwords like early fusion, BEV (Bird’s Eye-View), mapless and the ongoing end-to-end.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*qYCfd50MDykVQSohubLCZA.png" /><figcaption>Exhibit 3: Evolution of Deeproute solution, source: Deeproute.ai</figcaption></figure><p>DeepRoute claimed to be working with several local and international OEMs at the moment and the first batch of about 10,000 vehicles carrying its end-to-end systems would hit the market within the year. Although not officially confirmed, one of the OEMs is said to be Great Wall Motor.</p><h3><strong>SenseAuto (SenseTime)</strong></h3><p>SenseTime has started devoting efforts in the automotive sector since 2017 (reference <a href="https://medium.com/geekculture/chinese-computer-vision-startups-riding-the-auto-wave-27cc536d997d">post</a>). It officially introduced the brand name SenseAuto in 2021 aiming to become an automotive Tier1 supplier. Till now, it has delivered smart cabin solutions to major OEMs covering 100+ vehicle models. However, the attempts on smart driving didn’t get much attention until the new wave of end-to-end autonomous driving, where the company’s specialties in AI are crucial.</p><p>Obviously the team got the idea of end-to-end in 2017 when Honda requested an intelligent driving solution with cameras only (no HD map). A research team at SenseTime followed the direction ever since and came up with UniAD in late 2022 that won CVPR 2023 Best Paper Award. UniAD looks like a design of modular end-to-end system (reference <a href="https://medium.com/@schen583/chinas-4-stage-approach-to-end-to-end-autonomous-driving-b018b3c706ec">post</a> “Stage 3”). It coordinates five essential tasks in a planning-oriented philosophy. The query design serves as interfaces connecting all nodes resulting in flexible intermediate representations and exchanging multi-task knowledge toward planning (exhibit 4).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*tPhRhtIYi9IZwhKaVG2gLw.png" /><figcaption>Exhibit 4: Pipeline of UniAD, source: arxiv.org/pdf/2212.10156</figcaption></figure><p>In recent months, SenseAuto demonstrated the capabilities of UniAD system on complex city roads with cameras only. In terms of model training, the company currently collects data from test drive and obtains synthetic data through world model. After the system is deployed on mass production vehicles, which is expected in 2025, partner OEMs would feed the data back to SenseAuto. It has established a supercomputing center in Shanghai that currently boasts over 45,000 GPUs with a total computing power of 12,000 PFLOPS.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*V71G_5D4g9z4NCG1d89LUA.png" /><figcaption>Exhibit 5: Evolution to UniAD, source: SenseAuto</figcaption></figure><h3><strong>Pony.ai</strong></h3><p>Pony.ai is one of the leading L4 autonomous driving technology companies with growing fleets of robotaxi and robotruck across China. It claimed to have proceeded to one end-to-end model in August last year and the model has been updated onto L4 robotaxis and L2 mass production passenger cars.</p><p>Pony sees its advantage in end-to-end autonomous driving as multiple data sources for training that include robotaxi fleets, L2 passenger cars and roadside sensors. It has also built up evaluation systems and simulation software over the years to better process data.</p><p>The company has teamed up with ROX motor, an EV startup that launched the first model in 2023 with a few hundred sales volumes per month, and Pony supplied the intelligent driving system in the premium version. It has not disclosed any other projects with OEMs.</p><h3><strong>Horizon Robotics</strong></h3><p>Horizon Robotics is known for developing cost-effective AI chips used in passenger vehicles’ intelligent driving systems as well as providing software and hardware co-optimized solutions to OEMs and Tier1s. The company was reported to have consolidated the solution team in October of last year dedicating all efforts on a high-level intelligent driving solution, later revealed as SuperDrive. It is built on the J6P chip with 560 TOPS, the most advanced one among Journey 6 series launched earlier this year.</p><p>SuperDrive adopts a world model in the perception task with sensors and navigation map as inputs, which outputs both feature-level info and dynamic &amp; static objects. The feature representations then feed into an end-to-end planner and the object list feeds into a hybrid planner with some hand-coded rules and a validation network. Through an interactive evaluation process, the hybrid planner would eventually generate the optimal trajectory (exhibit 6).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*1alhB05xuGcySUdTt2MRdw.png" /><figcaption>Exhibit 6: SuperDrive architecture, source: Horizon Robotics</figcaption></figure><p>Horizon recently demonstrated SuperDrive performance on pubic roads. It claims to be working with a number of OEMs on delivering the solution and the first vehicle model is expected to hit the market in Q3 2025.</p><h3><strong>PhiGent Robotics</strong></h3><p>PhiGent Robotics was founded in 2021 aiming to become an intelligent driving solution provider. The company is fairly young while the co-founders come with extensive experience in the semiconductor industry previously with such as AMD, Horizon and Baidu, probably for which reason it has completed six funding rounds so far with the latest one being Pre-B of $30 million.</p><p>PhiGent is looking to differentiate with pure vision and stereovision, in other words affordable cost, enabling the highly demanded city NOA functions. It claims to have won multiple projects with PhiGo, a mid-level intelligent driving solution, of which the basic version is built with 7 cameras only on a single Horizon J6E chip.</p><p>Meanwhile, the company teams up with Geely research institute and has together proposed GraphAD for end-to-end autonomous driving algorithm that employs a graph model to describe the complex interactions in traffic scenes. GraphAD will be incorporated in its high-level intelligent driving solutions to be launched in 2025. The CTO of PhiGent believed that the boom of large language models in 2023 in fact inspired a path for end-to-end autonomous driving to go from concept to reality. Eventually, they will deploy a multimodality foundation model on the cloud to work together with onboard GraphAD.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*1Z04rcwIZfPCxR_2X1KnGw.png" /><figcaption>Exhibit 7: GraphAD intro slide, source: PhiGent Robotics</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*mEH3_95JT7gQgUH5yDHZVw.png" /><figcaption>Exhibit 8: GraphAD on the vehicle and foundation models on the cloud, source: PhiGent Robotics</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e545fa9380f1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Chinese Auto OEMs Embracing the End-to-end Paradigm]]></title>
            <link>https://schen583.medium.com/chinese-auto-oems-embracing-the-end-to-end-paradigm-78042c8d2438?source=rss-e5271e521780------2</link>
            <guid isPermaLink="false">https://medium.com/p/78042c8d2438</guid>
            <category><![CDATA[china]]></category>
            <category><![CDATA[self-driving-cars]]></category>
            <category><![CDATA[electric-vehicles]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[autonomous-driving]]></category>
            <dc:creator><![CDATA[Shuai Chen]]></dc:creator>
            <pubDate>Mon, 05 Aug 2024 09:52:03 GMT</pubDate>
            <atom:updated>2024-08-05T09:52:03.971Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/578/1*MgOYgwZBJzNIBIRs331_9A.jpeg" /></figure><p>The autonomous driving community in China has witnessed a paradigm shift towards an end-to-end, learning-based approach from the rule-based method that has been mainstream in the nearly two decades since the first DARPA Grand Challenge. In my last post, we discussed the background of this shift, the concept of end-to-end and the four development stages to realize one-model end-to-end. Now we look into the automotive OEMs in China.</p><h3><strong>Xpeng</strong></h3><p>On May 20th, Xpeng started pushing a new version of OTA updates (XOS 5.1.0) with an end-to-end model for intelligent driving. According to the CEO’s introduction at a press conference, it was the first end-to-end model on mass production vehicles in China, which also marked a new phase in the company’s city NOA development being mapless, with phase 1 being HD map reliance in 2022 and phase 2 being partially mapless.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*oiKhPbVNnJGCZQImXTQdRg.png" /><figcaption>Xpeng CEO introducing city NOA development phases</figcaption></figure><p>The model consists of three parts namely XNet, XPlanner, and XBrain. XNet is performing perception tasks with a deep neural network integrated with dynamic, static and vision-based 2K occupancy networks. XPlanner is doing planning and control through another AI model trained by massive 5-star drivers’ driving data, which was a key update as the planner used to be based on hand-coded rules.</p><p>XBrain is a large language model to do reasoning in order to understand and react properly to special lanes, road signs and behavioral commands for example. It works like the human brain to handle complex and unseen scenarios with safety in mind.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*ka7ZdWSwao5PPSwUJYeNgg.png" /><figcaption>Xpeng end-to-end model in three parts</figcaption></figure><p>Xpeng believed the current version would double its intelligent driving capabilities. There is an internal iteration every other day based on video training, on-road testing and simulated driving and the performance is expected to leap by 30 times in the next 18 months. The company has reserved 7000 of Nvidia A100 as computational power.</p><p>Xpeng believes the end-to-end system will eventually enable L4 and fully autonomous driving.</p><h3><strong>Li Auto</strong></h3><p>In July, Li Auto made OTA 6.0.0 available to all users that included a mapless NOA function, enabled by a system with two modules being perception and planning, each is a learning-based model. Meanwhile, it has revealed the next generation solution in details, which will be a one-model end-to-end system.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*8uGgWt33ByWv4s6oCend6Q.png" /><figcaption>Li current system in the middle and next gen one-model on the right</figcaption></figure><p>The team designed a two-system architecture with inspirations from the book ‘Thinking, Fast and Slow” by Daniel Kahneman. System 1 is an end-to-end framework dealing with majority of the driving scenarios, similar to human driver’s instinct based on experience and habits. System 2 is a large vision language model (VLM) to perform reasoning, analysis and calculation, dealing with the complex 5% of driving scenarios. Li plans to deploy the two systems on the vehicles, each on an Nvidia Orin since their current solution is built on two Orin chips. Additionally, there is a world model on the cloud to do verification and validation of system 1 and 2.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*BSLKoaP0cws1Q7RJ78nLfg.png" /><figcaption>Li to deploy E2E and VLM on the vehicles and world model on the cloud</figcaption></figure><p>The VLM model of system 2 has strong capabilities to understand complex driving environments like road surface, light conditions, and geo-specific traffic rules to aid system 1 in the decision-making process. It has 2.2 billion parameters and in order to run the model on-board, the team has successfully optimized the reasoning time of VLM from 4.1 seconds to 0.3 seconds.</p><p>Li expects the new solution to be available by the end of this year or early next year.</p><h3><strong>Nio</strong></h3><p>In July, Nio promoted the OTA updates of Banyan 2.6.5 featuring an end-to-end AEB with the latest version trained by 10,000 collisions from 2 billion km driving data. However, the end-to-end AEB phrase is a bit controversial from a technical perspective, for probably which reason the company focused on introducing Nio World Model (NWM) during the annual innovation day.</p><p>Nio believes the core of achieving end-to-end would be building a world model with the capability to do reconstruction and reasoning, which means to understand 3D space and predict future events respectively. NWM was described as a multivariable autoregressive generative model that takes inputs from video clips, navigation information, vehicle ego-motion, commands, etc. and output predictions and eventually a driving trajectory. It currently has the ability to generate videos in 2 minutes length, which means to predict the driving scenes in the next 2 minutes. It can generate 216 possible trajectories per 0.1 second for the optimal decision.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*vL71dAvQiA4bXvW5HdtyRQ.png" /><figcaption>Nio World Model illustration</figcaption></figure><p>Large amount of data is crucial to continually train the model. Nio is leveraging crowdsourcing AI to obtain real driving data from consumer vehicles and an in-house simulator (NSim) to generate synthetic data. The team would also be working on implementing NWM onto vehicles to do some demonstrations in Q4 of this year.</p><p>Although the wordings vary, what Xpeng, Li and Nio have in common are an end-to-end perception module and a recently upgraded planning &amp; control module enabled by some kind of AI models. There are continuous efforts in lining up computational power, betting on fundamental principles behind model training and building data engine.</p><p>A number of OEMs have revealed their directions in end-to-end autonomous driving without much detail. For example, <strong>Xiaomi</strong> has adopted end-to-end models in perception and decision, and recently in its AVP (Auto Valet Parking) function. <strong>Jiyue</strong> (formally known as Jidu) <strong>Auto</strong> CEO announced that end-to-end would be the key direction of R&amp;D. <strong>IM Motor</strong> is fully devoted to deploying the end-to-end architecture with the Tier 1 partner Momenta.</p><p><strong>GAC Motor</strong> has established a team and made some progress in the end-to-end autonomous driving solution. <strong>Great Wall Motor</strong> was reported to be working with another Tier 1 partner besides the internal incubated Haomo.ai to implement end-to-end solutions.</p><p>In the next post, we will look at automotive Tier 1s.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=78042c8d2438" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>