In five days, Ant Group's Robbyant shipped four open releases that together cover the entire pipeline of physical intelligence — from how a robot sees, to how it acts, to how it imagines the world, to how it controls its own body in real time.
A self-supervised 1B Vision Transformer built for dense spatial perception, not just semantic labels. It leads NYU-Depth v2 RMSE — ahead of DINOv3 — and rivals or beats foundation models up to 7× larger on dense tasks.
Why it's the base layer: Robots don't fail because they can't name objects — they fail because they don't know precisely where surfaces and boundaries are in 3D space. Prioritizing geometry over semantics is the right foundation for everything stacked on top.
A 6B vision-language-action model that maps 20 different robot configurations into one shared action space, trained on robot trajectories plus 10,000+ hours of egocentric video. Built to cross the gap "from foundation to application" — precision manipulation, long-horizon execution, and cross-embodiment generalization.
Why it matters: The hardest problem in robotics isn't one smart robot — it's one model that works across every robot a company owns. Twenty embodiments in a single action space is a direct shot at that.
The only interactive world model to hit hour-level (infinite) generation without visual drift. A distilled real-time variant drives 720p at 60fps, and it introduces an agentic harness: a pilot agent plans character behavior while a director agent seeds new environment elements — turning a frame predictor into a self-sustaining world. Ships as a 14B model plus a 1.3B single-GPU version.
Why it matters: A robot that can imagine the near future can plan against it. Borrowing the "model + harness" pattern from coding agents (Codex-style scaffolding) and applying it to world modeling is the conceptual leap here.
An autoregressive diffusion model that predicts while acting — interleaving video and action tokens in one causal sequence via a Mixture-of-Transformers, then re-grounding on the real robot's latest observation every step. A partial-denoising trick plus an asynchronous inference pipeline lets it run high-frequency closed-loop control, outperforming SOTA policies (including π0.5) on long-horizon and precision manipulation.
Why it's the top layer: Open-loop policies drift because they roll out long segments without checking reality. Re-grounding on real observations at every step — while staying causal — is what separates a lab demo from a robot you can actually deploy.
Who's Actually Leading in Physical AI
Everyone claims "embodied AI." Here's who's actually shipping models the field is building on — ranked by what they've put in the open and on real hardware, not press releases.
Owns the substrate everyone else builds on — Cosmos world foundation models, the GB300 compute stack, and the sim-to-real pipeline via Isaac Lab. Doesn't need to win the robot race; it wins by powering all of them. CoreWeave, Alibaba Cloud, and dozens of robotics labs have integrated NVIDIA's full physical AI stack. Structurally the hardest player to displace — picks up every time someone trains or deploys a robot anywhere. |
Shipped four open models in five days — LingBot-Vision (1B param, spatial perception), LingBot-VLA 2.0 (vision-language-action), LingBot-World-Infinity (causal world model), and LingBot-VA 2.0 (video-action). The 1B vision model reportedly outperforms Meta's DINOv3 at a fraction of the scale. All open-weight, freely downloadable. The most aggressive open push in the field right now — and a signal China is betting on open models to lead the hardware wave. |
Gemini Robotics brings frontier vision-language reasoning directly into robot manipulation — think of it as ChatGPT for robot hands. The RT (Robotics Transformer) lineage set early cross-task benchmarks the field still measures against. Deepest research bench of any lab. Weakness: mostly closed, so the broader developer community can't build on it directly. Long-term this may slow them relative to open competitors. |
Founded in 2024 by Stanford's Chelsea Finn. The π0.5 and π0.7 models are the current reference cross-embodiment manipulation policies — trained on the largest robot interaction dataset to date, using co-training across heterogeneous tasks for open-world generalization. Most new VLA (Vision-Language-Action) research benchmarks against π0.x. Small headcount, outsized field influence. Released openpi on GitHub. Punching well above its size. |
The most accessible full-size humanoid on the market — the G1 starts at $16,000, the research-grade H1 at $90K. Both ship with open-source SDKs, ROS 2 support, and 3D LiDAR. Completing a 1.9km autonomous marathon course in April 2026 and a synchronized Kung Fu performance at the Spring Festival Gala with dozens of G1s, Unitree is proving reliability at scale. The robot of choice for researchers who want to ship, not wait. |
Leading on real-world deployment: Figure 03 is live in BMW's Spartanburg and Leipzig factories (12,000 units/year production capacity); Tesla Optimus targeting 50,000+ units at its own factories. Figure raised $675M from NVIDIA, Microsoft, Jeff Bezos, and OpenAI. Their edge is proprietary fleet-scale training data generated by deployed robots — a flywheel that gets better the more robots are running. Model stacks stay closed and vertically integrated. Winning the product, not the open ecosystem. |
Digit is the quiet proof-of-concept that has been hiding in plain sight: 75 units live inside Amazon's Spanaway, WA fulfillment center as of April 2026, having already moved over 100,000 totes. First humanoid robot company to go public ($2.5B deal). Less glamorous than a humanoid doing backflips — but moving 100K totes in a real warehouse is the most credible deployment data in the field. |
Rankings reflect model quality, open-source impact, and real-world deployment — not valuation or press coverage.