Much of the recent conversation about AI has been about software: chat assistants, coding agents, retrieval. In October 2026 the centre of gravity visibly shifted towards physical AI, models that perceive the real world, decide what to do and move a robot body to do it. Within a few days, Google said its new Gemini 4 Argon model had taught a humanoid a throwing game from a video, NEURA Robotics introduced a multisensory foundation model for its robot fleet, Galaxea published a vision-language-action model that reasons and acts with a single set of weights, and South Korea cleared 20 humanoids to work alongside people on factory floors.
This article explains the technology underneath these announcements, separates what has been demonstrated from what has only been claimed, and offers a practical way for technology leaders to evaluate physical AI for their own operations.
What a vision-language-action model is
The term vision-language-action (VLA) model comes from Google DeepMind's RT-2 paper (2023). The idea is simple: take a vision-language model that was pretrained on internet-scale images and text, and teach it to output robot actions as if they were just another language. RT-2 encoded each action, the movement and rotation of the robot's gripper plus how far to open it, as a short string of number tokens, and trained the model on robot demonstrations and web data together.
The result was a robot policy that inherited some of the web model's understanding. Across about 6,000 evaluation trials, RT-2 generalised much better than earlier models to objects, backgrounds and instructions it had never seen in robot data, for example placing an object near a particular number or icon. The paper was also candid about the limits that still matter today: the robot did not learn new motions from web data, because its physical skills were bounded by the robot demonstrations, and the largest 55-billion-parameter model ran at only 1–3 Hz from a cloud service.
The second building block is data from many robots. The Open X-Embodiment collaboration pooled more than one million real robot trajectories from 22 robot types and 21 institutions, including Stanford, UC Berkeley and Carnegie Mellon, covering 527 skills. Models trained on this mixture showed positive transfer: the larger RT-2-X performed roughly three times better than RT-2 on skills that existed only in another robot's data. That finding, that one model can learn from many bodies, is the foundation of every "one brain, many robots" claim you hear today.
What changed in 2026: five shifts
1. Robot foundation models are going beyond cameras
NEURA Robotics' Neuron AI, announced on 14 October, is described as a Perception-Language-Action model that combines 3D vision, touch and force, audio, robot state and language, with memory. NEURA's argument is practical: only sound and touch confirm that a lock has opened, and a robot must remember objects that have left its field of view. The company says the model is trained on real robot data, simulation-amplified data and human demonstrations captured with sensor suits, and runs across five of its robot embodiments. It frames progress around a business metric, Time-to-Skill: how quickly a robot can learn a new task and perform it reliably.
2. Reasoning is moving inside the action stream
Galaxea's G0.5, published on 15 October, puts chain-of-thought reasoning and action generation into a single autoregressive stream. The model first writes out subtasks, target-object bounding boxes and 2D trajectory hints, then generates action tokens, replanning at every frame. Because instructions enter the same stream, language can shape behaviour at run time: in one example, adding "push harder" made the robot apply enough force to switch on a toaster. Galaxea reports an 82.5% average success rate across 10 tabletop tasks in a zero-shot DROID evaluation and stronger results than π0.5 on the long-horizon BEHAVIOR-1K tasks, although the weights have not yet been released.
3. Video and simulation are becoming the training ground
Google's Gemini 4 Argon launch on 13 October included an internal robotics test: the model watched a raw video of a person playing cornhole, mapped the environment, computed whole-body joint movements and wrote the simulation code itself. Transferred to Apptronik's Apollo humanoid, the robot sank seven consecutive shots in a real-world trial. Google calls this a "Video-to-Sim-to-Real" pipeline and says it compresses weeks of manual simulation modelling into a few hours. Google also reports an 87% score for Argon on the ERQA embodied-reasoning benchmark.
The open ecosystem is moving in the same direction. In January, NVIDIA released Cosmos world models for synthetic data generation and policy evaluation in simulation, Isaac GR00T N1.6, an open reasoning VLA for humanoid full-body control, and Isaac Lab-Arena for large-scale robot policy evaluation, with GR00T models available through Hugging Face's LeRobot library.
4. Hardware is being designed for learning
Boston Dynamics' new hand for Atlas shows how AI now shapes mechanical design. The hand has four fingers and 13 degrees of freedom, up from 7, dense pressure sensors on the fingertips and palm, and backdrivable, directly actuated joints. It was built to be simulated with high fidelity so that reinforcement learning in simulation transfers to hardware. The team dropped the pinky after deciding the extra dexterity was not worth three more actuators, with their cost, volume and failure risk. Boston Dynamics also makes a point every buyer should hear: wearable data-collection devices capture useful context, but not the exact action signals needed to control a dexterous hand, which is why the company sees reinforcement learning in simulation as essential.
5. Regulation is starting to catch up
On 13 October, South Korea's regulatory sandbox committee granted an exemption that lets 20 bipedal humanoids from Robros work in spaces shared with people at a semiconductor packaging firm, an auto-parts maker, a shipyard and a logistics company. According to the Seoul Economic Daily, previous rules treated industrial robots as stationary equipment behind safety fences; the exemption is conditional on risk assessments and safety measures, and the data collected will inform future safety standards.
Reading the announcements like an executive
As a technology executive, I find these developments genuinely significant, and I also read them with discipline:
- Most numbers are self-reported. The cornhole run is described by Google as an internal test; G0.5's benchmarks come from Galaxea's own report. Treat them as strong signals of direction, not as guarantees for your site.
- Demos are not deployments. An eleven-minute laundry sequence, which NEURA plans to show at NVIDIA GTC Berlin, proves coordination over a long task. It does not tell you the intervention rate across a thousand shifts.
- The RT-2 limits have not disappeared. Models still generalise best within the skills their robot data covers, and control frequency still constrains what can run in the cloud versus on the robot.
- Data is the moat. Every serious player now talks about its data pipeline as much as its model: real robots, simulation and human demonstrations. Your own processes and skilled workers are a data asset.
How to evaluate physical AI for your operation
- Start from the process, not the robot. List tasks that change often, require handling varied parts or tools, or are dangerous. Fixed, high-volume tasks are usually still better served by classic automation.
- Define the metrics before the pilot. Time-to-Skill, task success under realistic variation, human interventions per hour, cycle time and recovery from failure. A single success video is not a metric.
- Ask about the data contract. Who owns demonstrations and fine-tuned skills? NEURA, for example, says customers can keep their skills private. Make this explicit in every agreement.
- Insist on simulation and evaluation tooling. Policy changes should be tested in simulation and on a benchmark suite before reaching the floor, the way we test software releases.
- Plan the compute. Decide what runs on the robot and what runs on servers. Edge modules are getting more capable; NVIDIA prices its Jetson T4000 at $1,999 at 1,000-unit volume within a 70-watt envelope.
- Treat safety as a design input. Shared-space operation needs risk assessments, clear stop and override procedures and logging of every action, exactly the conditions Korea attached to its exemption.
- Keep architecture options open. Open VLA models and frameworks lower the cost of experimentation and reduce dependence on a single vendor.
Closing thought
Vision-language-action models are doing for robots what large language models did for software: turning many narrow, hand-programmed systems into one adaptable foundation. The difference is that mistakes in the physical world are measured in broken parts and safety incidents, not wrong answers. The organisations that benefit first will not be those that buy the most impressive demo, but those that build the data, evaluation and safety discipline to turn robot foundation models into reliable work.
References
- Gemini 4 Argon: Built to solve your hardest problems, Google, 13 October 2026
- NEURA Robotics Introduces Neuron AI: The Brain of the Physical World, NEURA Robotics, 14 October 2026
- Galaxea G0.5: A New-Generation Embodied Foundation Model That Enables Robots to Think While Acting, Galaxea Dynamics, 15 October 2026
- Robot Hands for Modern AI and Real Work, Boston Dynamics
- NVIDIA Releases New Physical AI Models as Global Partners Unveil Next-Generation Robots, NVIDIA Newsroom, 5 January 2026
- Korea Clears Safety Rules to Put 20 Humanoid Robots on Factory Floors, Seoul Economic Daily, 14 October 2026
- RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control, Brohan et al., Google DeepMind, arXiv:2307.15818
- Open X-Embodiment: Robotic Learning Datasets and RT-X Models, Open X-Embodiment Collaboration, arXiv:2310.08864