<?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[Black Coffee Robotics - Medium]]></title>
        <description><![CDATA[Autonomous Robots: we use state of the art algorithms and tools such as ROS/ROS2, simulation environments and development pipelines to provide robotics solutions in domains such as mobile robots (indoor and outdoors), drones, arm manipulators and boats! - Medium]]></description>
        <link>https://medium.com/black-coffee-robotics?source=rss----27efe4fc31f3---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Black Coffee Robotics - Medium</title>
            <link>https://medium.com/black-coffee-robotics?source=rss----27efe4fc31f3---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 00:20:20 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/black-coffee-robotics" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[MoveIt Motion Planning: How OMPL, CHOMP, and PILZ Planners Actually Behave on ROS2]]></title>
            <link>https://medium.com/black-coffee-robotics/moveit-motion-planning-how-ompl-chomp-and-pilz-planners-actually-behave-on-ros2-19cefcf421c0?source=rss----27efe4fc31f3---4</link>
            <guid isPermaLink="false">https://medium.com/p/19cefcf421c0</guid>
            <category><![CDATA[motion-planning]]></category>
            <category><![CDATA[moveit2]]></category>
            <category><![CDATA[robotics]]></category>
            <dc:creator><![CDATA[Gaurav Gupta]]></dc:creator>
            <pubDate>Wed, 05 Aug 2026 22:56:16 GMT</pubDate>
            <atom:updated>2026-08-05T22:56:16.704Z</atom:updated>
            <content:encoded><![CDATA[<p><a href="https://moveit.picknik.ai/main/index.html">MoveIt2</a> is the de facto framework for <a href="https://docs.ros.org/en/humble/index.html">ROS2</a> motion planning on robotic arms. Most teams reach it through the move_group node and an RViz panel, and never look closely at the planner that turns a goal pose into a joint trajectory. That planner is where the reliability, speed, and path-quality tradeoffs live. This article stays on the planning algorithms MoveIt exposes and the pipeline around them, on ROS2 Humble; perception, grasping, and full task orchestration are out of scope except where they exist to cover a planner&#39;s weakness. It continues a theme we have written about before, in <a href="https://www.blackcoffeerobotics.com/blog/curobo-nvidia-and-ros2-for-motion-planning">cuRobo vs MoveIt2</a> and <a href="https://www.blackcoffeerobotics.com/blog/ros2-moveit-servo-a-tool-for-robot-teleoperation-and-data-collection">ROS2 MoveIt Servo</a>.‍</p><h3>Three branches of planners, three different cores</h3><p>MoveIt exposes several planner families that solve the problem in fundamentally different ways. Knowing which branch you are on explains most of the behavior you will see.‍</p><p><strong>Sampling-based (OMPL).</strong> The default. Planners in the <a href="https://ompl.kavrakilab.org/">Open Motion Planning Library</a> randomly sample collision-free joint configurations and connect them into a tree until start and goal link. They are probabilistically complete but say nothing about repeatability. <a href="https://ompl.kavrakilab.org/classompl_1_1geometric_1_1RRTConnect.html">RRTConnect</a> grows two <a href="https://en.wikipedia.org/wiki/Rapidly_exploring_random_tree">Rapidly-exploring Random Trees</a> and returns when they meet: fast, but jagged and non-deterministic run to run. <a href="https://ompl.kavrakilab.org/classompl_1_1geometric_1_1RRTstar.html">RRT*</a> is asymptotically optimal, rewiring to shorten the path for as long as it is allowed to run, which means it uses its entire time budget.‍</p><p><strong>Trajectory optimization.</strong> <a href="https://www.ri.cmu.edu/pub_files/2009/5/icra09-chomp.pdf">CHOMP</a> deforms an initial trajectory by covariant gradient descent over a distance field; <a href="https://www.researchgate.net/publication/221078155_STOMP_Stochastic_trajectory_optimization_for_motion_planning">STOMP</a> does the same with gradient-free sampling. Their cost landscape is non-convex, so a poor initial guess lands in an in-collision local minimum and the planner fails. STOMP’s MoveIt2 plugin is not available on ROS2 Humble.‍</p><p><strong>Analytical (PILZ).</strong> The <a href="https://moveit.picknik.ai/main/doc/examples/pilz_industrial_motion_planner/pilz_industrial_motion_planner.html">PILZ Industrial Motion Planner</a> deterministically generates industrial PTP, LIN, and CIRC motions from kinematic limits. It does not plan around clutter, so we describe it but leave it out of the benchmark: in these obstacle scenes it would refuse or drive into an obstacle, and its behavior is known by construction.‍</p><h3>Why your MoveIt2 arm moves badly when the plan looks fine</h3><p>The planner is one stage in a pipeline that reshapes its output. move_group owns the <a href="https://moveit.picknik.ai/main/doc/examples/planning_scene/planning_scene_tutorial.html">planning scene</a> (robot model plus perception), runs pre-planning adapters to fix the start state, calls the planner plugin, then runs post-planning adapters that time-parameterize the path with TOTG or <a href="https://github.com/pantor/ruckig">Ruckig</a> for jerk-limited smoothing. The key model: OMPL returns a path, not a trajectory. Velocity, acceleration, and jerk limits are added afterward, so much of &quot;the arm moves badly&quot; is an adapter problem, not a planner problem.</p><p>Configuration decides the rest. The most consequential OMPL parameter is allowed_planning_time on the request: a ceiling RRTConnect rarely hits, but effectively the runtime for RRT*. The CHOMP planner is tuned separately through ridge_factor, learning_rate, and clearance parameters that determine whether it converges or stalls in a local minimum.</p><p>‍</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5NWQGemLdYQSahJD.png" /><figcaption>Moveit2 Planning Pipeline</figcaption></figure><p>‍</p><h3>What breaks first: benchmarks on ROS2 Humble</h3><p>We benchmarked RRTConnect, RRTStar, and CHOMP on ROS2 Humble across a <strong>table use case</strong> (a pick query), a <strong>box</strong> scene (easy and hard pairs), and a <strong>cage</strong> scene with tight obstacles, over three runs each with a 10s budget, measuring planning time and joint-space path length (Jlen, in radians).</p><p>‍</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*b31hiO4ynbDLRrXWEFF4aQ.png" /><figcaption>Motion Plans using RRTConnect</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*3RCziJriMVd2zh5R.gif" /><figcaption>Motion Plans using RRT*</figcaption></figure><p>‍‍</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*W33lNDx0oY_zAOTB.png" /><figcaption>Motion Planner Benchmark: <em>Small sample (3 runs per scene), 10s planning budget. The table use case reports the pick query only. Numbers are averaged over successful runs; “n/a” means no successful plan was produced.</em></figcaption></figure><p>Three patterns matter:</p><ul><li><strong><em>RRT* always spends its full budget.</em></strong> Its planning time sits near 10s in every scene because it keeps optimizing until told to stop. It often returns the shortest Jlen, but you pay the full budget every time.</li><li><strong>CHOMP failed in every scene.</strong> Seeded with a naive guess, it fell into in-collision local minima, which is why it is normally paired with a sampling-based pre-seed rather than run standalone.</li><li><strong>RRTConnect is the most reliable but wildly variable.</strong> It solved box-hard in 0.10s and dropped to 33% on box-easy: a start state near a collision boundary can make an easy query fail while a harder one succeeds.</li></ul><p>The real takeaway is variance, and hardware widens it. Raw OMPL paths are jerky without Ruckig smoothing, non-repeatability erodes trust in human-centric cells, and end-effector constraints rely on rejection sampling that balloons planning time.‍</p><h3>Making MoveIt planners production-grade</h3><p>Motion planning using MoveIt gives you broad planner coverage but weak guarantees. OMPL is complete but non-deterministic and slow under constraints, trajectory optimizers are smooth but fragile without a seed, and PILZ is deterministic but not collision-aware. None is production-grade out of the box. The engineering work is the layer around the planner: start-state hygiene, planner choice per query, jerk-limited time parameterization, constraint-aware seeding, and task-level sequencing with <a href="https://moveit.picknik.ai/main/doc/examples/moveit_task_constructor/moveit_task_constructor_tutorial.html">MoveIt Task Constructor</a>. Where CPU planning cannot keep up, GPU approaches like <a href="https://www.blackcoffeerobotics.com/blog/curobo-nvidia-and-ros2-for-motion-planning">cuRobo</a> are the next step.</p><p>That gap between a tutorial planner and a manipulation stack that runs reliably in a real cell is the work we do at Black Coffee Robotics. If you are hitting the reliability, constraint, or path-quality walls here, <a href="https://www.blackcoffeerobotics.com/">reach out to us</a>. We build and deploy robotics in the real world.‍</p><h3>Media &amp; source credits</h3><ul><li>MoveIt2, OMPL, CHOMP, STOMP, PILZ, and Ruckig documentation and papers are linked inline and belong to their respective authors.</li><li>Benchmark figures and scene visuals are from BCR’s internal RRTConnect / RRTStar / CHOMP runs on ROS2 Humble across the box, table, and cage scenes.</li></ul><p>‍</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=19cefcf421c0" width="1" height="1" alt=""><hr><p><a href="https://medium.com/black-coffee-robotics/moveit-motion-planning-how-ompl-chomp-and-pilz-planners-actually-behave-on-ros2-19cefcf421c0">MoveIt Motion Planning: How OMPL, CHOMP, and PILZ Planners Actually Behave on ROS2</a> was originally published in <a href="https://medium.com/black-coffee-robotics">Black Coffee Robotics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[URDF Viewer: A robotic workcell analysis and visualization tool]]></title>
            <link>https://medium.com/black-coffee-robotics/urdf-viewer-a-robotic-workcell-analysis-and-visualization-tool-9193edda9baa?source=rss----27efe4fc31f3---4</link>
            <guid isPermaLink="false">https://medium.com/p/9193edda9baa</guid>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[motion-planning]]></category>
            <category><![CDATA[workspace]]></category>
            <dc:creator><![CDATA[Gaurav Gupta]]></dc:creator>
            <pubDate>Wed, 04 Mar 2026 10:56:01 GMT</pubDate>
            <atom:updated>2026-03-04T10:56:00.962Z</atom:updated>
            <content:encoded><![CDATA[<p>‍</p><p>URDF Viewer is a powerful, browser-based platform designed for robot workcell analysis, workspace visualization, and feasibility assessment. Whether you’re a robotics engineer, system integrator, or sales professional, URDF Viewer can help you answer one critical question before any complex programming begins:</p><p><strong>Can this robot physically accomplish the required tasks within this specific workspace configuration?</strong></p><p>‍</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/998/0*IFLrJGlZa86JvIiy.png" /><figcaption>Urdfviewer user interfce</figcaption></figure><h3>‍The problem</h3><p>When evaluating whether a robot can handle a specific application, discussions typically focus on:</p><ul><li><strong>Technical specifications</strong> (payload capacity, reach radius, speed)</li><li><strong>Robot capabilities</strong> (degrees of freedom, repeatability, accuracy)</li><li><strong>Software algorithms</strong> (motion planning, path optimization, collision avoidance)</li></ul><p>While these factors are important, there’s a critical element that’s frequently overlooked: <strong>the physical configuration of the robot workcell itself</strong>.‍</p><p>Based on our experience in robotics deployment, we’ve discovered that many challenging robotics problems can be solved more efficiently by optimizing the physical setup before diving into complex algorithmic solutions. This includes carefully considering:</p><ul><li><strong>Robot base placement</strong> and mounting position</li><li><strong>Obstacle positioning</strong> within the workspace</li><li><strong>Task region location</strong> and orientation</li><li><strong>Approach angle constraints</strong> for specific operations</li><li><strong>Clearance requirements</strong> for robot motion‍</li></ul><h3>The current way to approach this</h3><p>When evaluating robot workcell feasibility today, professionals typically face two choices.</p><p>The first option involves setting up a full ROS/MoveIt environment, which requires installing frameworks, importing robot models, and building collision scenes. While powerful for final implementation, in this approach:</p><ul><li>Setup takes days or weeks</li><li>Demands specialized technical expertise</li><li>Making quick iteration is slow‍</li></ul><p>The second option is purchasing vendor-specific simulation software, which comes with its own challenges:</p><ul><li>License costs</li><li>Requires training and learning curves</li><li>Limited to specific robot brands with no platform flexibility</li></ul><p>Both approaches are good at what they’re designed for, but they create barriers when you simply need a quick feasibility check.‍</p><h3>Our Solution: URDF Viewer</h3><p>We built <strong>URDF Viewer</strong> specifically to bridge this gap as a web-based platform that works with any robot, allows custom workspace setup, requires no download, and has a quick learning curve.</p><p>The goal is simple:<strong> workcell feasibility and robot placement decisions in minutes, not weeks.</strong></p><p>Core design principles behind URDF viewer are:</p><ul><li><strong>Robot Compatibility: </strong>Work with any robot that has a URDF model not limited to specific brands or vendors.</li><li><strong>Zero Installation Required: </strong>Runs entirely in your web browser. No downloads, no installations, no IT approvals needed.</li><li><strong>Minimal Learning Curve: </strong>Intuitive interface designed for quick onboarding. Start evaluating workcells within minutes, not days.</li><li><strong>Rapid Iteration: </strong>Test multiple workspace configurations quickly. Adjust robot placement, move obstacles, redefine task regions, and immediately see the impact.</li><li><strong>Early-Stage Decision Support: </strong>Get critical feasibility answers in minutes, enabling informed decisions about robot selection and workspace design before significant resources are committed.</li></ul><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FmTnApgz62lA%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DmTnApgz62lA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FmTnApgz62lA%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/018effedb986a4ca04caa0d7477e6abf/href">https://medium.com/media/018effedb986a4ca04caa0d7477e6abf/href</a></iframe><h3>What you can do with URDF Viewer?</h3><h4>1. Select from wide range of robot models:</h4><p>Quickly compare different robots in the same workspace without switching platforms or dealing with vendor-specific tools. Test whether a cobot or industrial robot is better suited for your application with just a few clicks.</p><p>‍</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*HPpO7mW80RC6lYJm.png" /><figcaption>Explore multiple robot models in URDF Viewer</figcaption></figure><p>‍</p><h4>2. Move the robot around interactively:</h4><p>URDF Viewer offers multiple intuitive ways to control and position robots in your workspace. Use visual gizmo controls for drag-and-drop end-effector positioning with real-time feedback, enter precise pose coordinates for exact spatial positioning, or adjust individual joint angles to understand configuration constraints. This flexibility lets you quickly explore workspace possibilities or validate specific target positions, adapting to whether you’re brainstorming layouts or testing known requirements.‍‍</p><h4>3. Insert obstacles and define task regions:</h4><p>Create realistic workcell environments through simple drag-and-drop interaction. Add and configure obstacles with simple clicks. Define task regions where robots need to operate. This interface enables rapid modeling of real-world workspace constraints.‍</p><h4>4. Set approach angle constraints and run workspace analysis:</h4><p>URDF Viewer leverages <strong>AI-powered analysis</strong> to evaluate workcell feasibility comprehensively. Define regions with approach angle constraints for specific tasks and run detailed assessments that identify which task regions are reachable, visualize failures with color-coded displays, and explain why positions fail in natural language.‍</p><h3>Who is this for?</h3><p>Our goal is to help professionals and enthusiasts across robotics. Sales engineers gain quick feasibility assessments during customer meetings and faster sales cycles. System integrators and manufacturing engineers streamline workcell layout design and robot placement optimization while reducing engineering time. Students and hobbyists access professional-grade learning tools for free, exploring robot kinematics interactively without complex setup.‍</p><h3>Try URDF Viewer Today</h3><p>URDF Viewer is live and ready to use at <a href="https://urdfviewer.com/">urdfviewer.com</a>. No account registration, downloads, or installations required, just open your browser and start analyzing workcells immediately.</p><p>Have a specific workcell scenario you’re trying to solve? Need help evaluating a particular robot configuration or workspace layout? We’d love to <a href="https://www.blackcoffeerobotics.com/">connect</a> with you and discuss your unique challenge.</p><p>‍</p><p>‍</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9193edda9baa" width="1" height="1" alt=""><hr><p><a href="https://medium.com/black-coffee-robotics/urdf-viewer-a-robotic-workcell-analysis-and-visualization-tool-9193edda9baa">URDF Viewer: A robotic workcell analysis and visualization tool</a> was originally published in <a href="https://medium.com/black-coffee-robotics">Black Coffee Robotics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Robot Palletizing Software Requires More Than Basic Automation]]></title>
            <link>https://medium.com/black-coffee-robotics/why-robot-palletizing-software-requires-more-than-basic-automation-ed73e5198321?source=rss----27efe4fc31f3---4</link>
            <guid isPermaLink="false">https://medium.com/p/ed73e5198321</guid>
            <category><![CDATA[warehouse-management]]></category>
            <category><![CDATA[motion-planning]]></category>
            <category><![CDATA[pallets]]></category>
            <category><![CDATA[robotics]]></category>
            <dc:creator><![CDATA[Gaurav Gupta]]></dc:creator>
            <pubDate>Wed, 25 Feb 2026 09:32:24 GMT</pubDate>
            <atom:updated>2026-02-25T09:32:24.360Z</atom:updated>
            <content:encoded><![CDATA[<p>Robotic palletization appears deceptively simple on paper: pick boxes, stack them on pallets, repeat. Yet anyone who has attempted to deploy a palletizing system in a real production environment quickly discovers that the gap between “<strong>simple in theory</strong>” and “<strong>reliable in practice</strong>” is substantial.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Xw4cVY-8h9uFo0wJ.jpeg" /><figcaption>Illustration of a robotic palletizing process</figcaption></figure><p>‍After completing over 50 robotics projects across various industries, we’ve observed that palletization challenges rarely stem from the robot itself. The real complexity emerges from the interaction between workspace constraints, product variability, system integration requirements, and the need for consistent uptime.</p><p>This article examines critical areas where palletization projects typically encounter friction, and how proper engineering analysis during the planning and practices during the execution phase can prevent costly rework.‍</p><h3>Workspace Design and Motion Planning for Automated Palletizing Systems</h3><p>A common misconception is that if a location falls within a robot’s specified reach envelope, motion planning will work. In practice, kinematic feasibility is only the starting point. A robot palletization system might technically reach a position but lack sufficient joint range to approach from the required angle, or it might reach the position but be unable to maintain the straight-line trajectory necessary for stable box placement.</p><p>‍</p><h4>The Real-World Constraint</h4><p>Conveyor heights, pick zone locations, and pallet staging positions must all be optimized not just for individual reachability, but for motion efficiency across the entire operational envelope. A poorly positioned conveyor might force the robot into awkward configurations that add seconds to every cycle, reducing throughput by 20–30% compared to an optimized layout.</p><p>‍</p><h4>Our Approach to Workspace Analysis:‍</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/0*oh13DGXp0n5KECDg.png" /><figcaption>In-house tools for robotic workspace and motion analysis</figcaption></figure><p>‍We have seen and tackled reachability issues in multiple robotic applications. We address these constraints through computational workspace analysis during the design phase. Rather than relying on vendor datasheets or intuition, with our in-house tools, we simulate thousands of potential configurations across different robot models, evaluating not just reachability but motion efficiency, collision likelihood, and cycle time impact. This analysis is vendor-agnostic; we can evaluate solutions from <a href="https://www.fanucindia.com/">FANUC</a>, <a href="https://www.hd-hyundairobotics.com/en/main">Hyundai</a>, <a href="https://www.universal-robots.com/?utm_source=Google&amp;utm_source=Google&amp;utm_medium=paid-search&amp;utm_medium=paid-search&amp;utm_campaign=IN_SA_IN_General_always-on_paid-search_Google_2025&amp;utm_campaign=IN_SA_IN_General_always-on_paid-search_Google_2025&amp;utm_campaign_id=701P400000eNKNIIA4&amp;utm_campaign_id=701P400000eNKNIIA4&amp;utm_content=paid-ads-always-on&amp;utm_content=paid-ads-always-on&amp;gad_source=1&amp;gad_campaignid=23012364475&amp;gbraid=0AAAAADCsb0oCBtA_hA4Sy7Ye_lZElnnbB&amp;gclid=CjwKCAiA95fLBhBPEiwATXUsxDDFO457bylWEyQWokStmQ_OdP4UNhGSiv0V7QIip-Bil5ygZQPfOBoCDlwQAvD_BwE">Universal Robots</a>, or any other platform like <a href="https://robotiq.com/solutions/palletizing">Robotiq palletizer</a>, allowing equipment selection to be driven by actual performance data rather than existing vendor relationships.‍</p><h3>Robot and End-Effector Selection: The Throughput-Payload Tradeoff</h3><p>The choice between vacuum grippers and mechanical clamps isn’t just about holding force; it’s about how that choice cascades through robot selection, cycle time, and ultimately system economics.‍</p><h4>End-of-Arm Tooling Selection for Automated Palletizing</h4><p><strong>Vacuum systems</strong> offer flexibility across varying box sizes and surface finishes, but they introduce dependencies on compressed air quality and surface cleanliness.</p><p><strong>Mechanical clamps</strong> provide more deterministic grasping but require precise box positioning and may not accommodate significant dimensional variation.</p><p>The decision becomes more complex when considering multi-pick capabilities: gripping two boxes simultaneously can theoretically double throughput, but it also doubles the payload. It may require a larger robot with different kinematic characteristics.‍</p><h4>Why Standard Calculations Fall Short</h4><p>A collaborative robot rated for a 20kg payload might handle single-box palletising adequately, but adding a dual-pick gripper could push it beyond safe operating limits or slow joint velocities enough to eliminate any throughput gain. Conversely, oversizing to an industrial robot for future dual-pick capability might introduce unnecessary cost and footprint for the current application.</p><p>Proper selection requires dynamic simulation of the complete pick-place cycle, not just payload calculations. Joint velocities, accelerations, and trajectory planning behaviour all contribute to actual cycle time. We’ve seen cases where a lower-payload robot with optimised kinematics outperformed a higher-payload alternative simply due to better joint configuration for the specific workspace geometry. Each workstation is unique, and equipment selection should reflect actual operational requirements rather than general rules of thumb.‍</p><h3>System Integration</h3><p>Production environments rarely involve a single conveyor feeding a single pallet. Real deployments often require managing multiple input streams, different product lines converging on the same palletizing cell, or multiple pallets being built simultaneously to match order composition.</p><p>‍</p><h4>Beyond Mechanical Integration</h4><p>The integration challenge isn’t just mechanical. The system must coordinate:</p><ul><li>PLC communication and upstream buffering</li><li>Product tracking through vision systems</li><li>Pallet completion signalling and changeover timing</li><li>Synchronisation between multiple material flows</li></ul><p>A poorly designed integration can leave the robot sitting idle between motions, or in dual-robot scenarios, force the robots to work sequentially rather than operating in parallel through coordinated workspace sharing.</p><h3>Vision Integration for Error Detection and Recovery</h3><p>Vision-less palletization systems, which rely on mechanical guides and assumed box positions, can achieve impressive cycle times in controlled conditions. Their limitation isn’t speed; it’s brittleness. Without perception feedback, these systems cannot detect when something goes wrong until the error compounds into a more serious failure.‍</p><h4>The Cascading Error Problem‍</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ybvBYKD-GJf3CsB8.png" /><figcaption>Example Illustration of cascading palletization errors</figcaption></figure><p>‍A slightly misaligned box might be placed successfully once, but subsequent boxes stacked on that foundation inherit the error. By layer three or four, the entire pallet may be unstable, requiring manual intervention. Meanwhile, the robot continues operating normally because it has no way to detect the problem.‍</p><h4>Vision Systems: Implementation Matters</h4><p>Vision-guided systems introduce their own challenges:</p><ul><li>Camera positioning and lighting conditions</li><li>Computational latency and processing time</li><li>Pose estimation accuracy across product variations</li><li>Integration with existing control systems</li></ul><p>A poorly implemented vision system might be slower than a blind system without providing meaningful error detection.</p><p><a href="https://www.blackcoffeerobotics.com/blog/perception-in-robotics-from-geometry-to-generative-ai">Having worked with various camera systems and vision algorithms</a> across both indoor and outdoor scenarios, we design integration solutions specific to each facility’s needs</p><p>We give equal importance to the human machine interface: when the system detects an anomaly, operators need clear information about what happened and what intervention is required.‍</p><h3>Mixed-SKU Palletization</h3><p>Single-SKU palletization with fixed patterns is a well-understood problem. When every box is identical and arrives in a predictable sequence, pattern generation can happen offline, and the robot executes a predetermined choreography. If you are considering a Mixed-SKU palletization for your facility, this model will break entirely.‍</p><h4>The Runtime Decision Problem in Palletizing Software</h4><p>When box dimensions vary within a single pallet, placement becomes a real-time optimization problem. The system must decide where each box should go based on:</p><ul><li>What has already been placed</li><li>What is coming next in the queue</li><li>Stability and reachability constraints</li><li>Packing efficiency targets</li></ul><p>This requires runtime pattern generation, treating placement as a continuous decision problem rather than a fixed sequence‍</p><h4>From Theory to Practice</h4><p>Academic research on bin packing optimizes for density. Industrial palletization must balance density against cycle time, stability, and motion planning feasibility. A mathematically optimal packing might require approach angles the robot cannot achieve, or might create overhangs that violate stability constraints. The pattern generator must understand not just geometry, but the capabilities and limitations of the physical system.</p><h4>A Software-Centric Architecture‍</h4><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FHJ39oDzXxT4%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DHJ39oDzXxT4&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FHJ39oDzXxT4%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/9d5fe88d59f159ccb9d0622ba6db96a6/href">https://medium.com/media/9d5fe88d59f159ccb9d0622ba6db96a6/href</a></iframe><p>Here is a demonstration of a mixed-SKU palletization system that handles different SKUs with box dimensions varying by factors of two or more, without requiring manual reconfiguration between product changes.‍</p><h3>Moving Forward</h3><p>If you’re evaluating automated palletizing, we welcome a conversation about your specific requirements. Sometimes the right answer is off-the-shelf equipment with thoughtful integration. Sometimes it requires custom analysis to avoid costly mistakes. Either way, the implementation should start with understanding your operational constraints, not with a product catalogue.</p><p>We’re available to discuss workspace analysis, vision-guided palletizing<strong>, </strong>or whether your application might benefit from advanced mixed-SKU capabilities. <a href="http://blackcoffeerobotics.com/contact">Reach out</a> if you’d like to explore what’s possible for your facility.</p><p>‍</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ed73e5198321" width="1" height="1" alt=""><hr><p><a href="https://medium.com/black-coffee-robotics/why-robot-palletizing-software-requires-more-than-basic-automation-ed73e5198321">Why Robot Palletizing Software Requires More Than Basic Automation</a> was originally published in <a href="https://medium.com/black-coffee-robotics">Black Coffee Robotics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[2025 at Black Coffee Robotics, A Year in Review]]></title>
            <link>https://medium.com/black-coffee-robotics/2025-at-black-coffee-robotics-a-year-in-review-4fdae0fa4ac5?source=rss----27efe4fc31f3---4</link>
            <guid isPermaLink="false">https://medium.com/p/4fdae0fa4ac5</guid>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[motion-planning]]></category>
            <dc:creator><![CDATA[Gaurav Gupta]]></dc:creator>
            <pubDate>Wed, 07 Jan 2026 08:55:35 GMT</pubDate>
            <atom:updated>2026-01-07T08:55:34.680Z</atom:updated>
            <content:encoded><![CDATA[<p>2025 was a year of working in two modes at once. We were still building demos and exploring new ideas, but a much larger share of our time went into production readiness compared to previous years. Real hardware, real constraints, and systems that needed to run repeatedly and predictably shaped most of our work.</p><p>Across projects, demo quality and production quality were not separate phases, they had to evolve together. That tension defined the year.‍</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*c9S2EdkwVMyBuocJ.png" /><figcaption>Black Coffee Robotics: 2025 Recap</figcaption></figure><p>‍</p><h3>Project Work</h3><h4>Industrial robotic arms for wall panel construction</h4><p>A major part of the year was spent on an industrial robotic arms system for wall panel construction, using a dual-arm, rail-mounted setup. The work covered end-to-end system integration: industrial arms with limited ROS2 support, rail motion, perception-driven manipulation, and human-in-the-loop teleoperation for debugging, recovery, and iteration.</p><p>We deployed a motion planning stack, integrated perception tightly with motion execution, and pushed the system toward production-ready throughput. This was not a clean progression from demo to production. Both had to exist simultaneously, under schedule pressure. The team was stretched, but pulled through.</p><p>This project significantly deepened our experience with layout optimization, safety engineering, hardware integration, and millisecond-level motion planning, especially when vision and human feedback are part of the execution loop.</p><h4>Motion planning platform work with PickNik</h4><p>We started working with <a href="https://picknik.ai/">PickNik</a> this year around real-world usage of MoveIt Pro as a motion planning platform. The work has been focused on understanding how motion planning systems behave in actual industrial applications, beyond controlled demos or research setups.</p><p>This has given us strong exposure to deployment practices, application constraints, and the kinds of edge cases that determine whether a planning stack is usable in production. The collaboration has been a solid learning channel, and we are optimistic about building further in this direction.</p><h4>Mapping and performance on constrained hardware</h4><p>Another stream of work focused on productionizing mapping pipelines on budget-friendly hardware. The emphasis was on performance, particularly for floor mapping under tight compute and sensor constraints.</p><p>This reinforced a recurring pattern we see across projects: under real limits, careful systems engineering and profiling matter far more than algorithmic novelty.</p><h4>Lab automation with robotic arms</h4><p>We also worked on lab automation systems using robotic arms, collaborating with a globally distributed hardware team. The focus has been on reliability, integration, and turning individual capabilities into a usable end-to-end workflow. This effort is expected to come together in early 2026.‍</p><h3>R&amp;D Themes Across the Year</h3><h4>Motion planning: core expertise, real progress</h4><p><a href="https://www.blackcoffeerobotics.com/blog/curobo-nvidia-and-ros2-for-motion-planning">Motion planning </a>has been central across almost everything we worked on. We are comfortable saying this is an area of deep expertise for us. In 2025, that expertise was pushed into harder territory: constrained planning, performance under tight timing budgets, interaction with perception noise, and predictable behavior in application contexts.</p><p>We have already solved several practical issues that commonly block deployment, particularly around constraint handling and making planning behavior stable and understandable. There are still hard problems left, but motion planning is one of the areas where we feel strongest going forward.</p><h4>Simulation: necessary, still unsolved</h4><p>Simulation remains necessary and frustrating. We spent time working with high-fidelity simulation, including <a href="https://www.youtube.com/watch?v=tWUn_7oSVbc">deformable bodies</a> and <a href="https://www.youtube.com/watch?v=K8ej1VOlMog">particle-based approaches</a>. These are useful for select use cases, but remain heavy, brittle, and expensive to maintain.</p><p>Across projects, the lesson was consistent: simulation creates value only when tightly coupled to planning, control, and real-world validation. On its own, it rarely does.</p><h4>VLAs and foundation-model robotics</h4><p>There has been a lot of buzz around Vision-Language-Action and other foundation-model approaches. <a href="https://www.blackcoffeerobotics.com/blog/vision-language-action-vla-models-llms-for-robots">We explored this space as well.</a> The demos are compelling, and the momentum is real, but practical deployment remains difficult.</p><p>Data collection, infrastructure, task-specific fine-tuning, evaluation, and reliability guarantees are still hard problems. The gap between impressive demos and dependable applications is real. This is an area we are watching closely, but not one we are betting production systems on yet.‍</p><h3>Open Source Work</h3><p>Open source continues to be one of the most rewarding parts of what we do, because usage is an honest signal.</p><ul><li><a href="https://github.com/blackcoffeerobotics/bcr_arm"><strong>BCR Arm</strong></a><br>We <a href="https://www.blackcoffeerobotics.com/blog/bcr-arm-an-open-source-robotic-arm-for-ros2-gazebo-and-isaac-sim">launched bcr_arm</a>, a simulated robotic arm platform supporting ROS2, and Gazebo+Isaac Sim. Something thatpeople are actively using and extending, shaped by real feedback rather than internal assumptions.</li><li><a href="https://github.com/blackcoffeerobotics/bcr_bot"><strong>BCR Bot</strong></a><br>We continued to see strong community interest and external usage on <a href="https://www.blackcoffeerobotics.com/blog/isaac-sim-photorealistic-rendering-for-next-gen-robot-development">bcr_bot</a>, helping validate our release!</li><li><a href="https://github.com/blackcoffeerobotics/vector_pursuit_controller"><strong>Vector pursuit controllers </strong></a><br>In addition to continue maintaince, ROS2 Jazzy porting, we also released binaries for <a href="https://www.blackcoffeerobotics.com/blog/vector-pursuit-vs-pure-pursuit-ros2-controller-plugins">our nav2 controller plugin</a>, further reducing friction towards using it for fine path tracking!</li><li><a href="https://github.com/blackcoffeerobotics/gazebo-ros-actor-plugin"><strong>ROS2 actor plugin ports</strong></a><br>We also ported our actor plugins from ROS1 to ROS2, supporting new age of development involving human actors and realistic simulations..</li></ul><p>Seeing people use what we build, without us pushing it, remains one of the clearest indicators of value.‍</p><h3>A Note on AI Coding</h3><p>AI-assisted coding has been genuinely helpful for us, but only when used correctly. It does not replace understanding. A black-box approach does not work.</p><p>When you broadly know what you are doing and can guide the system, the speedups are real. Used blindly, it produces brittle software and the illusion of progress. It is leverage, not a shortcut.‍</p><h3>Looking Ahead to 2026</h3><p>Going into 2026, our focus is clear: accelerating real-world robot deployments.</p><p>We are doubling down on robotic arms, reliable motion planning, and application readiness. The goal is not novelty or hype, but systems that survive deployment and behave predictably under real constraints.‍</p><p><em>Originally published at </em><a href="https://www.blackcoffeerobotics.com/blog/2025-at-black-coffee-robotics-a-year-in-review"><em>https://www.blackcoffeerobotics.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4fdae0fa4ac5" width="1" height="1" alt=""><hr><p><a href="https://medium.com/black-coffee-robotics/2025-at-black-coffee-robotics-a-year-in-review-4fdae0fa4ac5">2025 at Black Coffee Robotics, A Year in Review</a> was originally published in <a href="https://medium.com/black-coffee-robotics">Black Coffee Robotics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[bcr_arm: An open source simulated arm for ROS2, Gazebo and Isaac Sim]]></title>
            <link>https://medium.com/black-coffee-robotics/bcr-arm-an-open-source-simulated-arm-for-ros2-gazebo-and-isaac-sim-b48f127a3990?source=rss----27efe4fc31f3---4</link>
            <guid isPermaLink="false">https://medium.com/p/b48f127a3990</guid>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[simulator]]></category>
            <category><![CDATA[isaac-sim]]></category>
            <category><![CDATA[gazebo-simulator]]></category>
            <dc:creator><![CDATA[Gaurav Gupta]]></dc:creator>
            <pubDate>Wed, 16 Jul 2025 07:41:34 GMT</pubDate>
            <atom:updated>2025-07-16T07:41:34.756Z</atom:updated>
            <content:encoded><![CDATA[<p>There has been rapid change in Gazebo distros over the last 2 years, it is hard to find a robotic arm that out of the supported multiple Gazebo versions. This results in a heavy time investment to get basic resources up and running. The documentation is often unclear and continuously evolving. The transition from <a href="https://www.blackcoffeerobotics.com/blog/ros-1-to-ros-2-migration">Gazebo Classic to the modern Gazebo framework</a> has resulted in multiple long-term support (LTS) distributions: <a href="https://gazebosim.org/docs/fortress/install/">Fortress</a>, <a href="https://gazebosim.org/docs/garden/install">Garden</a>, <a href="https://gazebosim.org/docs/harmonic/install/">Harmonic</a>, each with subtle but critical differences in their APIs, plugin architecture, and ROS integration.</p><p><a href="https://developer.nvidia.com/isaac/sim">Isaac Sim</a>, NVIDIA’s simulation platform for AI-powered robotics, recently became open source as part of the Omniverse ecosystem. While this makes high-fidelity simulation more accessible, its rapidly changing APIs and Python-based extensions still present a learning curve for new developers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yB3lMS307Yw6Jhwm.png" /><figcaption>bcr_arm in Nvidia ISAAC Sim</figcaption></figure><p>The rapid release cycles of both ROS2 and Gazebo have created compatibility challenges, making it difficult to find a stable robotic arm simulation that works out of the box across various platforms.</p><p>We introduce <a href="http://github.com/blackcoffeerobotics/bcr_arm">bcr_arm</a>, an open-source platform designed to help solve these technical hurdles.‍</p><h3>What is it?</h3><p><a href="http://github.com/blackcoffeerobotics/bcr_arm">bcr_arm</a> is an open-source, 7-degree-of-freedom (DOF) robotic arm simulation platform. We specifically chose a 7-DOF configuration because of its kinematic redundancy, which supports more complex manipulation tasks. This extra degree of freedom allows for null-space motion, enabling the arm to reconfigure its joints to optimize for secondary objectives without changing the end-effector pose.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FWlseQHCf2PU%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DWlseQHCf2PU&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FWlseQHCf2PU%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/a59d89deb0d519d998b0c804598520fc/href">https://medium.com/media/a59d89deb0d519d998b0c804598520fc/href</a></iframe><p>‍The release provides a pre-configured environment for today’s development stacks:</p><ul><li><strong>ROS 2 Humble</strong> with Gazebo Fortress on Ubuntu 22.04</li><li><strong>ROS 2 Jazzy</strong> with Gazebo Harmonic on Ubuntu 24.04</li><li>NVIDIA <strong>Isaac Sim</strong> — to support the <a href="https://www.blackcoffeerobotics.com/blog/isaac-sim-photorealistic-rendering-for-next-gen-robot-development">development of photorealistic and AI-driven</a> manipulation applications.</li></ul><h3>‍Core Features</h3><ul><li>Fully pre-configured simulation environment</li><li>Support for MoveIt 2 motion planning</li><li>Integration with ros2_control and controller interfaces</li><li>Compatible with Isaac Sim Action Graph, allowing you to use Isaac workflows with ROS 2-based arm simulations.‍</li></ul><h3>Future Work</h3><p>We are committed to maintaining compatibility with future LTS releases of ROS2 and Gazebo. We are also finalizing a vacuum gripper end-effector, which will be released shortly to support a wider range of manipulation and pick and place tasks.‍</p><h3>Conclusion</h3><p>Our goal with bcr_arm is to provide a goto reference with regards to robotic manipulator simulation with newer LTS versions of Gazebo, and we welcome all contributions to it! We believe that by providing a well-documented, and feature-rich platform, we can help accelerate the development of the next generation of robotic applications.</p><p>Check it out at <a href="http://github.com/blackcoffeerobotics/bcr_arm/"><strong>github.com/blackcoffeerobotics/bcr_arm/</strong></a></p><p>If you‘re looking to create advanced and large simulations for your robotic applications with gz-sim, <a href="https://www.blackcoffeerobotics.com/">reach out to us</a>!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b48f127a3990" width="1" height="1" alt=""><hr><p><a href="https://medium.com/black-coffee-robotics/bcr-arm-an-open-source-simulated-arm-for-ros2-gazebo-and-isaac-sim-b48f127a3990">bcr_arm: An open source simulated arm for ROS2, Gazebo and Isaac Sim</a> was originally published in <a href="https://medium.com/black-coffee-robotics">Black Coffee Robotics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What I Look for When Hiring Robotics Engineers (From a Startup Founder)]]></title>
            <link>https://medium.com/black-coffee-robotics/what-i-look-for-when-hiring-robotics-engineers-from-a-startup-founder-ff0ea72b190b?source=rss----27efe4fc31f3---4</link>
            <guid isPermaLink="false">https://medium.com/p/ff0ea72b190b</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[careers]]></category>
            <category><![CDATA[internships]]></category>
            <dc:creator><![CDATA[Gaurav Gupta]]></dc:creator>
            <pubDate>Thu, 08 May 2025 22:35:36 GMT</pubDate>
            <atom:updated>2025-05-08T22:35:36.566Z</atom:updated>
            <content:encoded><![CDATA[<p>Over a year ago, I wrote a piece on <a href="https://www.blackcoffeerobotics.com/blog/how-to-become-a-robotics-engineer"><em>How to Become a Robotics Engineer</em></a>. Since then, I’ve reviewed hundreds of applications for robotics roles — and realized there’s often a big gap between what we look for and what candidates bring. Here’s what I’ve learned as a technical founder, and what we now focus on at Black Coffee Robotics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*WKp10IsC6rIOLXpK.jpeg" /><figcaption>Our team at Black Coffee Robotics</figcaption></figure><p>‍</p><h3>1. Long-Term Fit Over Short-Term Gains</h3><p><em>Compound interest is the most powerful force in the universe.</em></p><p>I often come across profiles of candidates looking to work for 2–3 months to “gain experience” before moving on to a master’s degree or another job. There’s nothing wrong with that — but it’s not what we’re looking for.</p><p>We invest significant time and energy in on-boarding and training. We’re not looking for a quick fix or someone who’s just looking for an LoR, and some lines on a resume. We’re looking for future partners — people who want to go deep into robotics and grow alongside us.</p><p>‍</p><h3>2. Initiative and Real Work &gt; Credentials</h3><p>We’ve hired candidates from a spectrum of engineering colleges — some of the best public and private universities, as well as lesser-known ones. The hits and misses have come from all kinds of backgrounds.</p><p>We don’t care about your institute or your GPA — we care about your work.</p><p>Ask yourself:</p><ul><li>What have you built?</li><li>What was your role and impact on those projects?</li><li>Is there any code we can see on GitHub?‍</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*DaiJa883UVdnqa6e.png" /><figcaption>Organization of a sample package — clearly listing out dependencies, build, run, and configure instructions</figcaption></figure><p>‍Not just a raw, undocumented script — but something clean, reproducible, and ideally, used by others. Our team, for example, has been maintaining multiple open-source projects — <a href="https://github.com/blackcoffeerobotics/bcr_bot">bcr_bot</a>, the <a href="https://github.com/blackcoffeerobotics/vector_pursuit_controller">vector_pursuit plugin</a>, and others to name a few. More than the complexity of the code, we care about the usefulness of the code.</p><p><strong>We don’t just deliver code — we deliver working solutions.</strong> So if your code isn’t readable or testable, it doesn’t help us. Show us your best work — document it, publish it, and share the link!</p><p>Another strong signal is initiative. At one of our previous workplaces, my co-founder (then colleague) Zubin noticed the team had no version control system in place. On his own time, he set up a GitLab instance for the entire org — not because anyone asked him to, but because the problem existed, and he decided to solve it. That mindset matters.‍</p><h3>3. Clarity Over Jargon</h3><p>A common pattern in interviews:</p><p>I ask someone about their project, and they jump straight into technical jargon — which ROS nodes they wrote, what planner they used, which transformer model they plugged in.</p><p>Instead, try this:</p><ul><li>What was the project about?</li><li>Why was it important?</li><li>How did you approach it?</li></ul><p><em>If you can’t explain it simply, you don’t understand it well enough.</em></p><p>That’s it. No buzzwords needed. Sure, it takes some practice. But in our weekly or client meetings, when we discuss individual modules and projects — we need to be able to communicate clearly and effectively. We need to explain our work to each other, and to our clients.</p><p>Communication isn’t optional — it’s critical. We need people who can think, talk, and build clearly — so we can move fast and build real systems.‍</p><h3>Final Thoughts</h3><p>While a lot of career guidance focuses on technical skills, the ability to communicate, to work in a team, and to think long-term are just as important.</p><p>We’re not hiring for credentials — we’re looking for long-term thinkers, clear communicators, and hands-on builders. If this resonates with you, check out our careers page or drop us a line — <a href="https://www.blackcoffeerobotics.com/careers">we’re always looking to meet sharp minds</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ff0ea72b190b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/black-coffee-robotics/what-i-look-for-when-hiring-robotics-engineers-from-a-startup-founder-ff0ea72b190b">What I Look for When Hiring Robotics Engineers (From a Startup Founder)</a> was originally published in <a href="https://medium.com/black-coffee-robotics">Black Coffee Robotics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Vision-Language-Action (VLA) Models: LLMs for robots]]></title>
            <link>https://medium.com/black-coffee-robotics/vision-language-action-vla-models-llms-for-robots-f60ba0b79579?source=rss----27efe4fc31f3---4</link>
            <guid isPermaLink="false">https://medium.com/p/f60ba0b79579</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[simulation]]></category>
            <category><![CDATA[vlm]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[robotics]]></category>
            <dc:creator><![CDATA[Gaurav Gupta]]></dc:creator>
            <pubDate>Tue, 22 Apr 2025 10:25:18 GMT</pubDate>
            <atom:updated>2025-04-22T10:25:18.199Z</atom:updated>
            <content:encoded><![CDATA[<p>In the ever-evolving world of robotics, making machines interact with their environments as naturally and intelligently as humans do remains a central challenge. At <strong>Black Coffee Robotics</strong>, we’ve been exploring how large AI models can bridge that gap, especially through the emerging field of <a href="https://arxiv.org/abs/2502.06851"><strong>Vision-Language-Action (VLA)</strong></a> systems.</p><p>These models bring us one step closer to robotic assistants that can understand user intent from just an image and a sentence eliminating the need for handcrafting control policies or manually programming each behavior.</p><p>In this post, we dive into how we tested and fine-tuned a powerful open-source model, <a href="https://github.com/OpenVLA/OpenVLA"><strong>OpenVLA-7B</strong></a>, for real-world robotic manipulation tasks, including pick-and-place, drawer opening, and tool usage.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*9fAnmqoXrbZBCmFU.png" /></figure><h3>Why Vision-Language-Action (VLA) Models?</h3><p>Traditional robotic learning pipelines — such as <a href="https://spinningup.openai.com/en/latest/spinningup/rl_intro.html">Reinforcement Learning (RL)</a> and <a href="https://imitation.readthedocs.io/en/latest/algorithms/bc.html">behavior cloning</a> often demand extensive domain-specific engineering, large datasets of robot-environment interactions, precisely shaped reward functions, and prolonged training using GPU resources. These constraints hinder generalization, scalability, and practical deployment in real-world settings.</p><p><strong>Vision-Language-Action (VLA) models</strong> offer a compelling alternative by leveraging the synergy of <a href="https://en.wikipedia.org/wiki/Multimodal_learning">multi-modal learning</a> to overcome these limitations. Here’s why they are becoming increasingly important in robotics:</p><ul><li><strong>Data Efficiency via Pretraining:</strong><br>VLA models can be pretrained on large-scale internet or simulation datasets using vision-language pairs (e.g., images and captions or instructional videos), reducing the reliance on task-specific robot data. This pretraining helps bootstrap capabilities that generalize well to downstream robotic tasks with minimal fine-tuning.</li><li><strong>Natural Instruction Following:</strong><br>By incorporating natural language as an input modality, VLA models can interpret and execute high-level instructions like <em>“pick up the red block”</em> or <em>“open the left drawer”</em> a paradigm shift from manually coded action policies or symbolic planning pipelines.</li><li><strong>Cross-Domain and Cross-Platform Generalization:</strong><br>The shared embedding space for vision and language enables a single model to transfer knowledge across tasks, objects, and robotic embodiments. This means that a model trained in one domain (e.g., simulation or a different robot) can still be effective in another with limited adaptation.</li><li><strong>Unified Policy Learning:</strong><br>Instead of decoupling perception, task understanding, and control into separate modules, VLA models learn an end-to-end policy that jointly reasons about the scene, the goal, and how to act. This holistic approach improves robustness and simplifies the system design.</li></ul><p>In short, VLA models present a promising foundation for building general-purpose robots capable of understanding and executing tasks in diverse and dynamic real-world environments making them a cornerstone of next-generation <a href="https://arxiv.org/html/2407.06886v1">embodied AI</a>.‍</p><h3>How We Applied OpenVLA</h3><p>We used <a href="https://github.com/OpenVLA/OpenVLA"><strong>OpenVLA-7B</strong></a>, a <a href="https://huggingface.co/transformers/">transformer-based</a> model with a vision encoder and action decoder that processes a <strong>224×224 RGB image</strong> and predicts a <strong>7-dimensional action vector</strong> (for 3D movement, rotation, and gripper control).</p><p>To evaluate its performance, we tested OpenVLA on three different environments with increasing complexity:</p><h4>Simulation Environments</h4><p><a href="https://mujoco.org/"><strong>MuJoCo</strong></a><strong> (OpenVLA default tasks):</strong> For initial evaluation and baseline results.</p><p><a href="https://pybullet.org/wordpress/"><strong>PyBullet</strong></a><strong> (Pick task):</strong></p><ul><li>Simple cube-picking using <a href="https://www.trossenrobotics.com/widowx-250">Widow-X</a> arm</li><li>CPU-friendly and easy to set up</li></ul><p><a href="https://isaac-sim.github.io/IsaacLab/main/index.html"><strong>Isaac Lab</strong></a><strong> (Dishwasher unloading):</strong></p><ul><li>High-fidelity physics and rendering</li><li>Parallel GPU-based data collection</li><li>Used <a href="https://www.kinovarobotics.com/product/gen3-robot">Kinova Gen3 7-DOF arm</a></li></ul><h4>Data Collection Strategy</h4><ul><li>Episodic collection using <a href="https://github.com/google-research/rlds"><strong>RLDS format</strong></a></li><li>Collected RGB image and matching action at each step</li><li>Tasks included:</li></ul><p><strong>Pick task:</strong> 30 steps per episode, 30 episodes</p><p><strong>Dishwasher task:</strong> 60 steps per episode, 50 episodes</p><ul><li>Control frequency: <strong>5 Hz</strong></li><li>All data collected via <strong>hand-designed motion policies</strong>, with injected noise for diversity and robustness‍</li></ul><h4>Fine-tuning</h4><ul><li>Used <a href="https://arxiv.org/abs/2106.09685"><strong>LoRA (Low-Rank Adaptation)</strong></a> for efficient model tuning</li><li>Each task fine-tuned for ~10k steps (~4–5 hours)</li><li>Trained on a single GPU with <strong>24 GB VRAM</strong>‍</li></ul><h3>Results &amp; Observations</h3><p>Our experiments yielded a blend of exciting breakthroughs and instructive limitations.</p><h4>Video Demonstration</h4><p>Below is a demonstration of OpenVLA-7B in action, showcasing its capabilities across different tasks:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8qSv7osJdTEHm41d.gif" /><figcaption><em>OpenVLA-7B demonstrating pick-and-place task.</em></figcaption></figure><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FzZ8BIjQYlaE%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DzZ8BIjQYlaE&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FzZ8BIjQYlaE%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/5806223d2283702cf6a165545efb4b2a/href">https://medium.com/media/5806223d2283702cf6a165545efb4b2a/href</a></iframe><h3>‍What Worked</h3><ul><li><strong>Simple pick-and-place tasks</strong>:<br>After fine-tuning, the model achieved near-perfect success rates (~100%) in picking up objects in uncluttered environments like PyBullet.</li><li><strong>Transferability</strong>:<br>The model adapted to different arms (Widow-X and Kinova) and environments with only modest task-specific data.</li><li><strong>Ease of Use</strong>:<br>The workflow — from data collection to finetuning — was relatively smooth and repeatable. Once the pipeline was in place, training new tasks was quick.‍</li></ul><h3>What Didn’t Work Well (Yet)</h3><ul><li><strong>Out-of-the-box performance</strong>:<br>Without fine-tuning, the pre-trained OpenVLA struggled with unseen environments (~0% success), reinforcing the importance of task adaptation.</li><li><strong>Complex, contact-heavy manipulation</strong>:<br>Tasks like <strong>unloading a plate from a dishwasher</strong> which involve object-object collisions, precise trajectory planning, and fine motor control — only achieved <strong>5–10% success</strong>, even after finetuning.</li></ul><p>‍</p><p>Here are some examples of failure cases we observed:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F1W_B00GDwA4%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D1W_B00GDwA4&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F1W_B00GDwA4%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/f897375ebf44b7293889707d100ea00c/href">https://medium.com/media/f897375ebf44b7293889707d100ea00c/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FvGrZaPPUOAc%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DvGrZaPPUOAc&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FvGrZaPPUOAc%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/26bb5a1c5f7af3c49cd1df47e8e398bc/href">https://medium.com/media/26bb5a1c5f7af3c49cd1df47e8e398bc/href</a></iframe><p>‍</p><ul><li><strong>Lack of long-horizon planning</strong>:<br>Since OpenVLA predicts only the <em>next step action</em>, it struggles in tasks where future planning is critical. This is where newer models like , which predict action sequences, may outperform.</li><li><strong>Data sensitivity</strong>:<br>We observed that <strong>the quality, diversity, and noise profile</strong> of the collected data dramatically affected learning. Adding variation helped, but too much noise or repetition degraded performance.‍</li></ul><h3>Key Takeaways &amp; Next Steps</h3><ul><li><strong>Action precision matters</strong>:<br>Fine manipulation is hard. Collision-aware actions and tactile feedback may be needed for complex tasks.</li><li><strong>Better data = better performance</strong>:<br>We’re now focusing on refining our <strong>data filtering and augmentation pipelines</strong>, ensuring the model sees the right kind of variety during training.</li></ul><h4><strong>Next-gen models are promising</strong></h4><p>We’re exploring <a href="https://www.physicalintelligence.company/blog/pi0"><strong>Pi0</strong></a> and similar models that:</p><ul><li>Accept multiple visual frames as input</li><li>Predict full action sequences instead of just one step</li></ul><h4><strong>Infrastructure is key</strong></h4><p>To scale experiments, we’re working on <strong>automated pipelines</strong> for data collection, visualization, and model debugging.‍</p><h3>Final Thoughts</h3><p>VLA models like OpenVLA are still maturing — but they already offer a glimpse into a future where robots can learn faster, adapt better, and interact more naturally. Our journey with OpenVLA showed both the promise and the path forward.</p><p>At <a href="https://www.blackcoffeerobotics.com/"><strong>Black Coffee Robotics</strong></a>, we’re excited to push the boundaries of intelligent robot control. If you’re exploring AI-driven robotics, manipulation, or custom automation solutions, reach out to us!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f60ba0b79579" width="1" height="1" alt=""><hr><p><a href="https://medium.com/black-coffee-robotics/vision-language-action-vla-models-llms-for-robots-f60ba0b79579">Vision-Language-Action (VLA) Models: LLMs for robots</a> was originally published in <a href="https://medium.com/black-coffee-robotics">Black Coffee Robotics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Isaac Sim: Photorealistic Rendering for Next-Gen Robot Development]]></title>
            <link>https://medium.com/black-coffee-robotics/isaac-sim-photorealistic-rendering-for-next-gen-robot-development-37ca8186c1d9?source=rss----27efe4fc31f3---4</link>
            <guid isPermaLink="false">https://medium.com/p/37ca8186c1d9</guid>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[simulation]]></category>
            <category><![CDATA[robot-operating-system]]></category>
            <category><![CDATA[open-source]]></category>
            <dc:creator><![CDATA[Gaurav Gupta]]></dc:creator>
            <pubDate>Wed, 13 Nov 2024 11:51:09 GMT</pubDate>
            <atom:updated>2024-11-13T11:51:09.599Z</atom:updated>
            <content:encoded><![CDATA[<p>The days of low-fidelity robot simulations are over. Modern autonomous robots heavily rely on visual perception through cameras and depth sensors. This makes photorealistic simulation environments critical for development and testing. NVIDIA’s <a href="https://developer.nvidia.com/isaac/sim">Isaac Sim</a> combines industry-leading RTX graphics technology with comprehensive robotics simulation capabilities to address this need.</p><p>‍In this article, we’ll explore how Isaac Sim enables photorealistic rendering for robotics applications and announce the support of <a href="https://github.com/blackcoffeerobotics/bcr_bot">bcr_bot</a> with Nvidia’s Isaac Sim — <a href="https://www.blackcoffeerobotics.com/blog/gazebo-simulator-migration-from-classic-to-ignition-with-ros2">our open-source mobile robot </a>simulation platform compatible with multiple ROS and <a href="https://gazebosim.org/home">Gazebo</a> distros.</p><h3>Why Photorealistic Rendering Matters</h3><p>Traditional robotics simulators like Gazebo focus primarily on physics simulation, with graphics being a secondary consideration. However, modern autonomous robots rely heavily on visual perception through:</p><ul><li>RGB cameras for object detection and semantic segmentation</li><li>Depth cameras for 3D scene understanding</li><li>LiDAR sensors for localization and mapping</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KeRzheMeUxy90XctJyZkng.png" /><figcaption>Side-by-Side Rendering Comparison: Gazebo vs. Isaac Sim Simulation.</figcaption></figure><p>For the perception algorithms to work reliably in the real world, the rendered environment needs to closely match reality — including proper lighting, shadows, reflections, and material properties.</p><h3>Isaac Sim’s RTX-Powered Rendering</h3><p>Isaac Sim leverages NVIDIA’s RTX technology and real-time ray tracing to deliver photorealistic rendering, featuring:</p><ul><li>Physically based materials and textures</li><li>Dynamic global illumination</li><li>Accurate shadows and ambient occlusion</li><li>Support for complex material properties like reflection and refraction‍</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8aRcsuA_J0YUSn9bkVnW2w.png" /><figcaption>Warehouse Demo in Isaac Sim: Realistic 3D Environment for Robotics Simulation.</figcaption></figure><p>This level of visual fidelity enables:</p><ul><li>Training Perception models on Synthetic data</li><li>Testing computer vision algorithms</li><li>Validating sensor fusion pipelines</li><li>Creating compelling demonstrations</li></ul><h3>Getting Started with bcr_bot in Isaac Sim</h3><p>To help developers get started with photorealistic robotics simulation, we’re excited to announce the support of <a href="https://github.com/blackcoffeerobotics/bcr_bot">bcr_bot</a> with Nvidia’s Isaac Sim</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FgfKOj_HZiX0%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DgfKOj_HZiX0&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FgfKOj_HZiX0%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/7f139395aa72b949e4768bfc04b0dc44/href">https://medium.com/media/7f139395aa72b949e4768bfc04b0dc44/href</a></iframe><p>‍<br>bcr_bot is now designed to serve as a reference implementation demonstrating Isaac Sim’s best practices around:</p><ul><li>Setting up sensors with realistic properties</li><li>Configuring materials for proper rendering</li><li>ROS2 communication</li><li>Scene composition and lighting</li><li>NAV2 integration</li></ul><p><strong>To get started:</strong></p><ol><li>Install Isaac Sim with ROS2 dependencies following the <a href="https://docs.omniverse.nvidia.com/isaacsim/latest/installation/index.html">official documentation</a>.</li><li>Clone the <a href="https://github.com/blackcoffeerobotics/bcr_bot">bcr_bot</a> repository</li><li>Follow setup instructions in the README to spawn bcr_bot in Isaac Sim.<br>‍</li></ol><h3>Looking Ahead</h3><p>We’re committed to maintaining bcr_bot as a community resource and plan to add:</p><ul><li>More sensor configurations</li><li>Additional demo environments</li><li>Integration with perception algorithms</li><li>Performance optimization guides</li></ul><p>‍</p><p>Are you looking to leverage photorealistic simulation for your robotics application? Reach out to us at <a href="https://www.blackcoffeerobotics.com/">Black Coffee Robotics</a>!</p><p>‍</p><p>‍</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=37ca8186c1d9" width="1" height="1" alt=""><hr><p><a href="https://medium.com/black-coffee-robotics/isaac-sim-photorealistic-rendering-for-next-gen-robot-development-37ca8186c1d9">Isaac Sim: Photorealistic Rendering for Next-Gen Robot Development</a> was originally published in <a href="https://medium.com/black-coffee-robotics">Black Coffee Robotics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Vector Pursuit: Controller Plugin for ROS2 Navigation]]></title>
            <link>https://medium.com/black-coffee-robotics/vector-pursuit-controller-plugin-for-ros2-navigation-07cd02d1bbb8?source=rss----27efe4fc31f3---4</link>
            <guid isPermaLink="false">https://medium.com/p/07cd02d1bbb8</guid>
            <category><![CDATA[autonomous-cars]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[navigation]]></category>
            <category><![CDATA[control]]></category>
            <category><![CDATA[robot-operating-system]]></category>
            <dc:creator><![CDATA[Gaurav Gupta]]></dc:creator>
            <pubDate>Sun, 18 Aug 2024 09:00:17 GMT</pubDate>
            <atom:updated>2024-08-18T09:00:17.381Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/729/0*1z-F0Ve8gVOYTzC-" /></figure><h3>Introduction</h3><p>Vector Pursuit is a geometric path-tracking algorithm that is based on the <a href="https://en.wikipedia.org/wiki/Screw_theory">Theory of Screws</a> and was first introduced in the paper: <a href="https://apps.dtic.mil/sti/pdfs/ADA468928.pdf">Vector Pursuit Path Tracking For Autonomous Ground Vehicles</a>. Vector Pursuit is used to track paths with a high degree of accuracy and relatively fast speeds while being computationally light and very simple to deploy. It simultaneously tracks both orientation and position on a path in a geometrically meaningful way.</p><p>The open-source ROS2 ecosystem contains several well regarded mobile robot controllers. ROS2 Navigation Stack (<a href="https://docs.nav2.org/">Nav2</a>) is the de-facto navigation software for a variety of robots. Keeping this in mind, we have engineered a <a href="https://github.com/blackcoffeerobotics/vector_pursuit_controller">Vector Pursuit</a> as a Nav2 plugin, so that it can be easily integrated with a range of ROS2 applications with ease. Our implementation outperforms existing Nav2 controllers in various metrics as discussed later in the article.</p><h3>Working Principle and Feature Offerings</h3><p>Vector Pursuit uses a <a href="https://www.mathworks.com/matlabcentral/answers/833328-what-is-lookahead-distance">look-ahead distance</a> to define a current goal pose on the path it is required to follow. It calculates the desired motion of the vehicle to reach this goal by considering both its location and orientation while remaining geometrically meaningful and computationally efficient. Vector Pursuit is based on Screw Theory, which describes the instantaneous motion of a rigid body relative to a coordinate system as a screw-like movement around a fixed axis. This controller is applicable to differential, legged and Ackermann mobile robots.</p><h4>Additional Features Offered</h4><p>In addition to the core path-tracking algorithm, the Vector Pursuit plugin also offers additional features which are described in brief. All these features can be toggled and tuned using <a href="https://github.com/blackcoffeerobotics/vector_pursuit_controller?tab=readme-ov-file#configuration">ROS2 parameters</a>.</p><ul><li><strong>Look-ahead distance scaling</strong>: The look-ahead distance is scaled proportionally to the robot’s current linear velocity.</li><li><strong>Collision awareness</strong>: The robot’s velocity is automatically lowered in proximity to obstacles and can be halted completely to avoid imminent collision.</li><li><strong>Goal overshoot prevention</strong>: To avoid overshoot, the controller will lower its approach velocity when nearing a final goal pose.</li><li><strong>On point rotation</strong>: When the trajectory to a look-ahead point deviates from the robot’s current heading by more than a certain angle, the controller will first rotate on point to avoid a large turning arc.</li><li><strong>Optional reversing</strong>: The controller is capable of generating reverse velocity if the path demands it.</li></ul><h3>Controller Comparison</h3><p>Two test setups were employed to conduct a comparison of some controllers offered by Nav2 against Vector Pursuit. One was a <a href="https://gazebosim.org/home">Gazebo</a> simulation and the other was a real-life, skid-steer robot. Certain controllers like Regulated Pure Pursuit and Dynamic Window Approach have comparable characteristics while others like MPPI are vastly different. The controllers tested against Vector Pursuit are described below.</p><ul><li><strong>Regulated Pure Pursuit (RPP)</strong> is used in path following scenarios with low compute resources. It is easy to tune but it does not take into account the orientation of the path and circumvent obstacles.</li><li><strong>Dynamic Window Approach (DWB)</strong> is used for path following scenarios where highly customizable behavior is required. Although it requires considerably more compute and tuning effort than RPP, it performs better for complex path-tracking and can circumvent obstacles.</li><li><strong>Model Predictive Path Integral (MPPI)</strong> is a path following controller which is ‘critic based’. It considers path orientation and has obstacle avoidance, which comes at a cost of high computation and considerable tuning.</li></ul><h4>Tracking A Sharp Turn At High Speeds</h4><p>The objective of this test is to compare the two most computationally efficient controllers Vector Pursuit and Regulated Pure Pursuit in terms of path-tracking quality. They are subjected to identical test conditions. In the figures below, the red line(composed of arrows) represents the odometry of the robot. The green line represents the desired path to be followed. In cases where the red line completely overlaps the green line the path is followed perfectly, deviations from the path will result in a visible green line.</p><ul><li>At <strong>1 m/s</strong>, Pure Pursuit begins to show slight deviations from the path and unwanted oscillations.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/489/0*3fSBIZraG5j9Gk3U.jpeg" /><figcaption>Pure Pursuit v/s Vector Pursuit path tracking at 1m/s.</figcaption></figure><ul><li>At <strong>1.5 m/s</strong>, Pure Pursuit shows complete failure in path-tracking with dramatic oscillations. Meanwhile, Vector Pursuit shows almost identical performance.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/489/0*BPRKYWfvKdIS0Shl.jpeg" /><figcaption>Pure Pursuit v/s Vector Pursuit path tracking at 1.5m/s</figcaption></figure><h4>Recorded Computation</h4><p>This test is meant to showcase the significant differences in CPU usage between the controllers. The Nav2 stack was run without <a href="https://docs.ros.org/en/foxy/Concepts/About-Composition.html">composition</a> on an Intel i5–1135G7 CPU at 2.40GHz with 16GB RAM. The CPU usage of the controller process was logged using the top command line utility and plotted using Python with a sliding window approach for smoothening. The same method was used to test the controllers on an ARM Cortex-A72 CPU @ 1.8 GHz with 8GB RAM.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FPNdqgSzP2yo%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DPNdqgSzP2yo&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FPNdqgSzP2yo%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/d6b7e49ff7c902810b0d3fa087fc8cc3/href">https://medium.com/media/d6b7e49ff7c902810b0d3fa087fc8cc3/href</a></iframe><p>‍</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FjYKILi296CU%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DjYKILi296CU&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FjYKILi296CU%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/6e42046004509d926cd07e05f8349ccc/href">https://medium.com/media/6e42046004509d926cd07e05f8349ccc/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*l3qpH82lmGGfw9A8.png" /><figcaption>CPU usage comparison from an arm64 RPI CPU on a real robot.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*odv7VhGRQfP9wvTh.png" /><figcaption>CPU usage comparison from an x86 Intel CPU with a simulated robot.</figcaption></figure><h4>Inferences</h4><p>As observed in the above experiments, Vector Pursuit is a robust replacement for Regulated Pure Pursuit (RPP) which struggles to provide reliable turning at high speed. Vector Pursuit displays comparable path-tracking performance to DWB while being simpler and easier to deploy and offers a much lower computational cost.</p><h3>Conclusion</h3><p>The release of the<a href="https://github.com/blackcoffeerobotics/vector_pursuit_controller"> Vector Pursuit Plugin</a> addresses the need for a computationally efficient, performant Nav2 controller. Leveraging the principle of Screw Theory, Vector Pursuit offers an optimal balance between precision and resource consumption, making it ideal for systems with limited computational power or scenarios demanding rapid response times.</p><p>We at <a href="https://www.blackcoffeerobotics.com/">Black Coffee Robotics</a>, are excited to offer this controller to the open-source robotics community and look forward to seeing how it enhances various autonomous navigation projects. The Vector Pursuit Controller is available on <a href="https://github.com/blackcoffeerobotics/vector_pursuit_controller">GitHub</a> as an industry ready controller plugin for Nav2 with 90% code coverage. We encourage developers to explore, utilize, and contribute to its continuous improvement.‍</p><p>Do you require optimized mobile robot control for your next robotics project? Reach out to us!</p><p>‍</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=07cd02d1bbb8" width="1" height="1" alt=""><hr><p><a href="https://medium.com/black-coffee-robotics/vector-pursuit-controller-plugin-for-ros2-navigation-07cd02d1bbb8">Vector Pursuit: Controller Plugin for ROS2 Navigation</a> was originally published in <a href="https://medium.com/black-coffee-robotics">Black Coffee Robotics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Motion Planning with Nvidia cuRobo and ROS2]]></title>
            <link>https://medium.com/black-coffee-robotics/motion-planning-with-nvidia-curobo-and-ros2-4b16fa6b27a9?source=rss----27efe4fc31f3---4</link>
            <guid isPermaLink="false">https://medium.com/p/4b16fa6b27a9</guid>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[robot-operating-system]]></category>
            <category><![CDATA[motion-planning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Gaurav Gupta]]></dc:creator>
            <pubDate>Mon, 29 Jul 2024 20:35:16 GMT</pubDate>
            <atom:updated>2024-07-29T20:35:16.623Z</atom:updated>
            <content:encoded><![CDATA[<p>Fast and reliable motion planning for robotic arms is essential. Tasks like packaging and palletizing objects of different shapes and sizes require high-speed, on-the-fly motion planning while precise maneuvers are necessary for applications like robotic surgery.</p><p>We at Black Coffee Robotics have seen an ever-increasing demand for faster motion planning capabilities to keep up with the rapidly evolving needs of the industry. In scenarios where traditional CPU-bound graph planners like <a href="https://en.wikipedia.org/wiki/Rapidly_exploring_random_tree">RRT</a> struggle to generate efficient motion plans and algorithms like <a href="https://github.com/tesseract-robotics/trajopt">Trajopt </a>often take several seconds to optimize trajectory, new solutions are needed. We tested one such solution i.e. Nvidia’s <a href="https://curobo.org/">cuRobo</a>, which reframes motion generation as a <a href="https://en.wikipedia.org/wiki/Global_optimization">global optimization problem</a>. By leveraging the power of GPUs, cuRobo claims to explore a vast search space of possible trajectories identifying optimal solutions in a fraction of the time required by conventional methods.‍</p><h3>A Conceptual Overview of Parallelization in Motion Planning</h3><p>The standard CPU-bound approach to motion generation using a graph planner involves a two-step process:</p><ol><li>Planning a single collision-free geometric path.</li><li>Smoothing that path for efficiency.</li></ol><p>In contrast, cuRobo uses parallel GPU processing to run many instances of collision-free <a href="https://en.wikipedia.org/wiki/Inverse_kinematics">inverse kinematics</a>(IK) to get multiple solutions. It then generates trajectory seeds by linearly interpolating between the start joint configuration and calculated IK solutions. Finally, it runs trajectory optimization on the seeds in parallel to obtain the best collision-free trajectory. This unlocks several key advantages, including improved trajectory quality and faster planning time.‍</p><h3>CuRobo: CUDA Accelerated Robot Library</h3><p>CuRobo leverages Nvidia’s parallel computing architecture to generate timestep-optimized, collision-free trajectories. We tested cuRobo because its Nvidia GPU-centric design allowed us to perform parallel motion planning on Nvidia GPUs, rather than relying on specialized hardware like FPGAs. Additionally, its integration with <a href="https://pytorch.org/">PyTorch</a>, offering a Python frontend, promises a more streamlined development process.‍</p><h3>Getting Started with cuRobo</h3><p>CuRobo has a simple setup process using NVIDIA’s robotics simulation platform <a href="https://curobo.org/get_started/2b_isaacsim_examples.html">Isaac Sim</a>. By following the <a href="https://curobo.org/get_started/2b_isaacsim_examples.html">official documentation</a>, you can configure and run cuRobo on any system that meets Isaac Sim’s minimum specifications. While an NVIDIA RTX graphics card is required to run Isaac Sim, cuRobo is compatible with non-RTX NVIDIA GPUs as well.</p><p>To remove cuRobo’s dependency on Isaac Sim for simulation and enable broader compatibility with any ROS2-supported simulator or hardware driver, we developed a custom ROS2 wrapper around cuRobo. This allows us to visualize motion plans in <a href="https://github.com/ros2/rviz/tree/humble">RViz</a> and provides a clear picture of the planned robot’s motion.</p><p>Here is the breakdown of the steps to get started with the cuRobo and ROS2:</p><ol><li><strong>Robot Configuration: </strong>CuRobo utilizes robot configuration files in addition to the standard URDF file. These configuration files are readily available for industrial robot arms on their <a href="https://github.com/NVlabs/curobo/tree/main/src/curobo/content/configs/robot">official Github</a>. and can be generated for custom robots following the provided<a href="https://curobo.org/tutorials/1_robot_configuration.html"> tutorial</a>.</li><li><strong>Development of ROS2 nodes: </strong>Develop<strong> </strong>ROS2 nodes to accept the desired goal pose for the robot arm and publish the trajectories generated by cuRobo based on the received goal pose.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fTjONyjwcPwYxDdr.gif" /><figcaption>Trajectory generated by cuRobo</figcaption></figure><h3>‍Comparison of cuRobo Motion Planning with MoveIt2 (RRTConnect)</h3><p>To conduct a head-to-head comparison between cuRobo and MoveIt2’s default planner, we tested motion generation for various goal poses using a simulated <a href="https://github.com/UniversalRobots">UR5 arm</a> on the same machine.</p><h4>Obstacle-Free Planning</h4><p>On testing various Cartesian poses within the arm’s workspace to generate motion paths, we found the following results.</p><ul><li><strong>CuRobo: </strong>Running cuRobo in obstacle-free environments resulted in a mean planning time of <strong>0.19 seconds</strong> and a success percentage of <strong>100%</strong>.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/880/0*M3bOgAU5gp6zsWBv.png" /><figcaption>CuRobo — no obstacles</figcaption></figure><p>‍</p><ul><li><strong>MoveIt2 (RRTConnect): </strong>Running MoveIt2 in obstacle-free environments resulted in a mean planning time of <strong>0.17 seconds</strong> and a success percentage of <strong>96%</strong>.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/880/0*_Olpy03cCtzaBnLp.png" /><figcaption>MoveIt2 — no obstacles</figcaption></figure><p><strong>Result: </strong>In our obstacle-free planning test, cuRobo demonstrated a slight edge in reliability and quality. CuRobo achieved a <strong>100%</strong> success rate and superior trajectory quality, despite a comparable mean planning time of <strong>0.19</strong> seconds to MoveIt2’s <strong>0.17 </strong>seconds.</p><p>‍</p><h4>Planning with obstacles</h4><p>On testing various Cartesian poses with obstacles in the workspace we observed the following.</p><ul><li><strong>CuRobo: </strong>Running cuRobo in an environment with obstacles resulted in a mean planning time of<strong> 0.69 seconds</strong> and a success percentage of <strong>90.32%.</strong></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/880/0*jfZAkPVcQsEpPJaL.png" /><figcaption>CuRobo with obstacles‍</figcaption></figure><p><strong>MoveIt2 (RRTConnect): </strong>Running MoveIt2 in the same environment with obstacles resulted in a mean planning time of <strong>0.33 seconds</strong> and a success percentage of<strong> 62.50%</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/880/0*Tyik0Sh_pQX9VVO6.png" /><figcaption>MoveIt2 with obstacles</figcaption></figure><p><strong>Result: </strong>In environments with obstacles, cuRobo demonstrated better reliability and trajectory quality. CuRobo achieved a <strong>90.32%</strong> success rate compared to MoveIt2’s <strong>62.50%</strong>, despite having a longer mean planning time of <strong>0.69</strong> seconds versus MoveIt’s <strong>0.33 </strong>seconds.‍</p><h3>Challenges and Limitations with cuRobo</h3><p>While cuRobo offers advantages, setting up cuRobo for custom robotic arms can be a tedious process. Configuring cuRobo for a non-supported robotic arm currently requires collision generation using Isaac Sim and a robot configuration file which is manually created using this <a href="https://curobo.org/tutorials/1_robot_configuration.html">template</a>. Ideally, the future should see more user-friendly, accessible tools or simplified workflows for configuring cuRobo with non-supported robotic arm setups without relying on Isaac Sim. Additionally, it’s important to note that the cuRobo <a href="https://github.com/NVlabs/curobo/blob/main/LICENSE">license</a> restricts its use to non-commercial purposes (with the exception of Nvidia Corporation and its affiliates). The only way to use it for commercial applications is to either ask Nvidia for a license, or use it via their <a href="https://github.com/NVIDIA-ISAAC-ROS/isaac_ros_cumotion">Moveit plugin</a>.</p><h3>‍Conclusion</h3><p>CuRobo is a compelling alternative for robot motion planning. It leverages parallelization on the GPU, leading to faster and higher-quality motion. Although custom arm setups currently have implementation challenges, cuRobo’s potential is evident. As development progresses, it could become a significant improvement for motion generation. Recently, NVIDIA released cuRobo as a <a href="https://github.com/NVIDIA-ISAAC-ROS/isaac_ros_cumotion">MoveIt2 plugin</a>, but it remains to be tested and is outside the scope of this article.‍</p><p>Are you planning to use accelerated motion planning for your next-gen robotic project? We’re here to help! <a href="https://www.blackcoffeerobotics.com/contact">Reach out to us</a>.</p><p>‍</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4b16fa6b27a9" width="1" height="1" alt=""><hr><p><a href="https://medium.com/black-coffee-robotics/motion-planning-with-nvidia-curobo-and-ros2-4b16fa6b27a9">Motion Planning with Nvidia cuRobo and ROS2</a> was originally published in <a href="https://medium.com/black-coffee-robotics">Black Coffee Robotics</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>