The Next “Mind”
By Ayush Gandhi | TKS Innovate NYC ’26
The Problem of Different Robot Bodies
A robot can learn to pick up a cup, another can learn to fold a shirt, and a humanoid can learn to walk across a room. But those skills do not automatically transfer between them.
The main reason is that every robot has a different physical structure, sensors, actuators, and range of movement. A machine with a simple gripper is fundamentally different from a humanoid with two hands, a torso, and dozens of joints. Those differences shape not just what a robot can do, but how it must execute each task, and therefore means that each robot must consume enormous amounts of data to thrive in the same fashion.
In robotics, this is called embodiment, the idea that the physical body through which an AI interacts with the world is dependent on its program. If one robot learns to grab an object by moving a certain joint, another robot might not even have that same joint and is thus limited in that distinct capability.. For one AI to control many different robots, it cannot just memorize its movements, but must have the “mental” capacity to understand that function and be able to understand how to procure that exact meaning using what it has.
Instead of thinking, “move joint 3,” it needs to understand the concept of something, similar to how we don’t look at someone giving a fist bump and say “close finger joints and extend elbow joint”, but rather, “give a fistbump”.
The distinction between understanding a task and reproducing a movement is the central hurdle to general-purpose robotics and an omnipotent AI mind.. A cup can be grasped by a two-finger gripper or a five-finger hand, but the motor commands are not interchangeable and therefore must be independently learned by each individual robot. An efficient reusable robotic intelligence must connect the shared goal to the abilities and limitations of each body. The real question is whether a single learned model can do that without being retrained from scratch for every single task.
Why a General Robot Brain Matters
If robots could transfer knowledge between different bodies, engineers would not have to train every robot completely from scratch and would instead be able to use a shared model across multiple systems. A warehouse robot, factory arm, research robot, or humanoid could potentially share the same basic intelligence while still doing different jobs without having to individually consume inefficient and environmentally unsustainable levels of data.
This could make robots much more useful in manufacturing, warehouses, research, and eventually everyday environments.
Today, robotic learning depends heavily on demonstrations. Sequences of camera images, joint states, sensor readings, and successful or unsuccessful actions are used to train individual robots.. Engineers might teleoperate a machine to show it how to move an object, then use recordings to train its policy. Those demonstrations differ from simply writing code as robots learn through access to VLA pattern which require a lot of hands on data gathering which can be a strenuous and annoying process to repeat for a wide variety of robots.. A shared model might reduce repeated data collection, but it would not eliminate the need for hardware engineering, testing, and safety supervision.
Training also carries a computational cost for training data. Epoch AI’s historical analysis of 124 notable machine-learning systems from 2009 to 2022 estimated that the dollar cost of compute for their final training runs grew approximately 0.51 orders of magnitude per year under one hardware price-performance method.
FIGURE 1
Historical estimated compute costs for notable ML training runs. The logarithmic vertical axis and regression uncertainty must not be interpreted as robot-specific costs. Source: Ben Cottier, Epoch AI (2023), “Trends in the dollar training cost of machine learning systems.”
How Vision-Language-Action Models Work
Google DeepMind and several competitors are exploring Vision-Language-Action models, or VLAs as a means of training this omnipotent mind of AI. These systems are not independent of LLM. Many develop upon existing widespread language-model and vision-language-model architectures, then extend them to predict physical actions. Instead of generating only text, a VLA visually processes information and stores it for future recognition.
The QUAR-VLA diagram below offers a concrete example. In this research system for quadruped robots, images and written instructions pass through feature-processing and tokenization steps before the model predicts action parameters.
FIGURE 2
QUAR-VLA / QUART combines visual observations and language instructions to generate quadruped control actions. This is an architecture example, not a comparison of different robot embodiments. Source: Ding et al., “QUAR-VLA: Vision-Language-Action Model for Quadruped Robots” (2023), arXiv:2312.14457.
This difference between task-level language and physical-level control is essential. The instruction “walk toward the door” conveys a goal for the robot to achieve, but a quadruped still needs to interpret and assess the cost of commands. A Robot arm instead needs appropriate joint or end-effector movements. Researchers are still trying to understand the formula to allow these systems to understand and develop systems that can recognize rather than “memorize”. This is where “embodiment” of an AI brain becomes difficult.
How Gemini Robotics Is Trying to Solve It
Google DeepMind is working on this problem through its Gemini Robotics program, which uses Vision-Language-Action models. These models combine what a robot sees, what a person tells it to do, and information about its physical capabilities.
Google, through DeepMind, is also developing Motion Transfer, which allows useful movement knowledge to transfer between different robot bodies. The intention behind this project is for AI to be able to transfer useful knowledge between robots, hence the idea of an “AI brain.”
That idea requires technical precision. Different machines have different action spaces which are defined as the number of the sets of numerical commands they can execute. In essence, a joint extension might not represent the same action requests from different robots. Cross-embodiment transfer means reusing task knowledge despite these differences, often by training on diverse robots and then adapting a model to a new target body. This is the idea of this article, replacing individual “training”, with holistic learning.
One of the strongest examples is Gemini Robotics On-Device 2. DeepMind has reported that the system can adapt to completely new two-arm robots with typically fewer than 200 examples, even when the new robot has different sensors, shapes, and degrees of freedom.
Google describes hours of adaptation to new embodiments in its testing for this idea. I have provided 2 examples of google robotic softwares, SO101 and Dexmate, which were used for post-training evaluations. The X axis of the original plot measures average hours of data per task, while the Y axis measures task success. Each point describes performance under the particular circumstances.
FIGURE 3
Original Google data-scaling plots for SO101 and Dexmate. GRODv2 consistently improves with added task data and ends at 53.3% and 75.6% success, respectively. Error bars reflect uncertainty. Source: Google DeepMind, “Gemini Robotics On-Device 2 Model Card” (July 2026).
At the largest plotted data budget, On-Device 2 reached 53.3% success on SO101 versus 6.7% for On-Device 1. On Dexmate, its final success was 75.6% compared with 33.3% for the older model. The results matter because they show improved sample efficiency. It showed that a pretrained model can have success, 53.33% to be exact, on performing a task on an untrained software. However, the data also warns against consistent success of this.
DeepMind is also working on embodied reasoning, where AI does more than simply move but can interpret a goal, form a plan, observe results, and adjust future actions.
The Race to Build a Universal Robot Brain
DeepMind is not alone. Several organizations are building competing systems for robot intelligence, and they do not all define the problem in exactly the same way nor approach the problem with teh same goa. Some focus on learning from many machines and then using those to run more, some emphasize continuous action generation in which robots focus on smoothing continuous signals, and others invest heavily in simulated demonstrations such as Google’s SO101.
Physical Intelligence: π₀
Purpose — generalist robotics policy for dexterous manipulation for robotic movement.
Scale — Trained on 7 robot configurations and 68 complex tasks
Method — Use VL understanding combined and mixed with flow-matching continuous movement.
Get Ayush Gandhi’s stories in your inbox
Join Medium for free to get updates from this writer.
Training — Relies on shared pretraining refined through task-specific post-training.
NVIDIA: GR00T
Purpose — GR00T N1 combines vision-language reasoning with action generation
Data USED — Uses a mix of real robotic demonstrations and simulations.
Training — Achieved 76.8% real-world success on GR-1 tasks, outperforming a 46.4% Diffusion Policy baseline.
Open X-Embodiment: Sharing the Training Data
Purpose — A collaborative project to solve the training-data bottleneck.
Scale — Collected over 1 million robot trajectories across 22 embodiments and hundreds of skills to lessen the training quality gap.
Goal- Makes diverse robotic training data reusable across different laboratories.
These 3 examples show the diverse nature of the robotics field and show how there is a lot of competition along this field. It shows and displays the diverse nature of innovation in this field, specifically pertaining to the development of an AI mind, but also displays advancement in other ways.
FIGURE 4
Composition of Open X-Embodiment. Panels distinguish dataset counts, scene counts, trajectory counts, common skills, and object labels. The corpus is varied but unevenly distributed. Source: Open X-Embodiment Collaboration, “Robotic Learning Datasets and RT-X Models” (2023/2024), Figure 2.
Taken together, these efforts suggest that a universal robot brain may require several breakthroughs at once. Physical Intelligence develops a shared action policy, NVIDIA explores powerful models paired with synthetic experience, Open X-Embodiment offers pooled robot data, and DeepMind emphasizes transfer and planning. While all these ideas may compete economically, their progress and innovation is shared in this one field.
What Is Still Holding Robots Back
Even with this progress, one AI cannot simply control every robot yet. A major problem is out-of-distribution situations, where a robot encounters something very different from its training.
A robot might work well in a clean room with familiar objects but struggle in a room that contains clutter, bad lighting, unfamiliar objects, or people moving nearby. Essentially the issue is not that a model cannot find an object in an internal database but that taken out of context it is harder for the model to understand and comprehend. It is that patterns learned in training can be unreliable when the physical environment changes and everything is not perfect. More varied training may help, but it does not guarantee success in every unfamiliar condition.
Speed and safety are also huge problems. A chatbot can take a second to respond without much consequence. A robot may need to react immediately if a person or object suddenly moves into its path. Because robots operate in the physical world, a mistake can damage equipment or hurt someone, which raises major concerns about reliability.
DeepMind has already addressed one part of this issue: On-Device 2 is built to run locally rather than depending on an internet connection for every action. That reduces network issues, but cannot eliminate sensing delays, motor constraints, unpredictable contact forces, and model mistakes that make this concept highly volatile and extremely unpredictable.
The robotics spectrum shown below is useful as a conceptual map of the ways and issues robots come across, not as a quantitative substantiated study.
FIGURE 5
Illustrative spectrum of robotics deployment and research. It expresses categories and challenges, not measured market share or a validated performance benchmark. Source: User-provided “Robotics Reality Spectrum” infographic; original author/source should be credited before publication.
A useful metric for future systems must therefore combine success, transfer, and safety. One impressive demonstration is not enough. A credible general model should complete tasks repeatedly, recover from unexpected conditions, work on several new robot platforms, and identify when it should stop or request intervention. This is the only way it can reach a level where it can properly adapt to humanistic needs and reach a point of independence.
Could One AI Eventually Control Everything?
I think robotics is moving toward a future where one general AI could control many different robot bodies. It may work somewhat like an operating system, with many different parts, but one overarching intelligence or brain that dictates movements and actions.
The difficulty lies not only in teaching the robot about existing systems but also in developing a system where AI can learn without enormous amounts of new data. That is what makes this technology exciting but also a little scary. If one intelligence can eventually control many different machines, improvements could spread across entire categories of robots instead of just one device. The possibilities could be almost unlimited, but so could the consequences if the systems are not reliable or safe.
My hypothesis is that by the year 2030, a leading research system will adapt one pretrained model to at least five previously unseen robot configurations, achieve at least 90% success across a shared set of manipulation tasks, and use fewer than 100 additional demonstrations per robot. This is just what I, Ayush Gandhi, not a robot or company, believes will happen based on my research and the growth I have seen.
Ultimately, the next “mind” may be defined not by how many tasks one machine can memorize, but by how much of its learning can be transferred to a body it has never encountered before. It will tested by the bounds on which an object can transfer its skills to a different enviorment and succeed in teh environment its tested in and the one it is challnged with. This is the future, that is without a doubt.
Sources and Image Credits
Epoch AI — Ben Cottier, “Trends in the dollar training cost of machine learning systems” (2023): https://epoch.ai/publications/trends-in-the-dollar-training-cost-of-machine-learning-systems
Ding et al., “QUAR-VLA: Vision-Language-Action Model for Quadruped Robots” (2023): https://arxiv.org/abs/2312.14457
Google DeepMind, “Gemini Robotics On-Device 2 Model Card” (2026): https://deepmind.google/models/model-cards/gemini-robotics-on-device-2/
Google DeepMind, “Gemini Robotics On-Device”: https://deepmind.google/models/gemini-robotics/on-device/
Google DeepMind, “Gemini Robotics Brings AI into the Physical World”: https://deepmind.google/blog/gemini-robotics-brings-ai-into-the-physical-world/
Google DeepMind, “Gemini Robotics 2 Brings Whole-Body Intelligence to Robots”: https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/
Google DeepMind, “Gemini Robotics: Embodied Reasoning”: https://deepmind.google/models/gemini-robotics/embodied-reasoning/
Google DeepMind, “Gemini Robotics ER 2 Model Card”: https://deepmind.google/models/model-cards/gemini-robotics-er-2/
Physical Intelligence, “π₀: A Vision-Language-Action Flow Model for General Robot Control”: https://www.physicalintelligence.company/download/pi0.pdf
NVIDIA, “Accelerate Generalist Humanoid Robot Development with NVIDIA Isaac GR00T N1”: https://developer.nvidia.com/blog/accelerate-generalist-humanoid-robot-development-with-nvidia-isaac-gr00t-n1/
Open X-Embodiment Collaboration, “Robotic Learning Datasets and RT-X Models”: https://arxiv.org/abs/2310.08864
Figure 5: User-supplied infographic; original creator/source not independently verified.
