Robot foundation models tracker: RT-2, GR00T, Gemini Robotics, dated claims and access

Robot foundation models are AI systems trained on broad data to control physical machines, and two of the biggest releases in the field landed within a day of each other this week. This tracker logs the major vision-language-action (VLA) and world action models, what each lab claims, what has been independently checked, and whether anyone outside the lab can run it. We update it whenever a new release or a correction changes the picture. The last update was November 12, 2026.

How to read this tracker

A VLA model takes camera images and a written instruction and outputs motor actions. The term was introduced in a July 2023 paper from Google DeepMind. A newer variant, which NVIDIA calls a world action model, predicts future video and robot actions together. Every entry below is dated and labeled by who ran the evaluation. "Vendor-reported" means the lab measured its own model on its own tasks. "Independent" means someone outside the lab reproduced it. At the time of writing, none of the entries below has an independent replication that we could find.

The log, oldest to newest

July 28, 2023: RT-2 (Google DeepMind). The paper "RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control" fine-tuned vision-language models on both robot trajectories and internet-scale tasks such as visual question answering, encoding robot actions as text tokens so they share a format with language. The authors report results from 6,000 evaluation trials and say the model generalized better to novel objects and followed commands absent from the robot training data. This is the paper that named the VLA class. Status: published preprint, vendor-reported.

March 18, 2025: GR00T N1 (NVIDIA). The paper describes "an open foundation model for humanoid robots" with a dual-system design: a vision-language module that interprets the scene and a diffusion transformer that generates actions in real time, trained jointly on real-robot trajectories, human videos and synthetic data. The authors report outperforming imitation learning baselines on standard simulation benchmarks and deploy it on the Fourier GR-1 humanoid. Status: preprint, vendor-reported, described as open.

March 25, 2025: Gemini Robotics (Google DeepMind). A VLA built on Gemini 2.0 that controls robots directly, plus Gemini Robotics-ER for embodied reasoning. The abstract says it can learn new short-horizon tasks from as few as 100 demonstrations and adapt to new robot bodies. Status: preprint, vendor-reported, limited access.

November 9, 2026: GR00T 2 (NVIDIA). An open world action model announced at the Conference on Robot Learning in Austin. NVIDIA reports 71% average success across 50 tasks from one demonstration per task and 95% from five, and trained on 100,000 hours of human first-person video that it is releasing publicly. Code is listed as available in December and no license is stated. The Next Web reports NVIDIA's paper says the tasks were chosen from ones the model could already learn well. Status: whitepaper published, code not yet released, vendor-reported.

November 10, 2026: Gemini Robotics 2.5 (Google DeepMind). DeepMind's most capable VLA to date, evaluated on more than 200 tasks, with success rates 30% to 100% higher than earlier baselines depending on task type and tasks finished 20% to 30% faster. A proof of concept at Mercedes-Benz sites in Kecskemét, Hungary, and Sindelfingen, Germany, reported a 96% success rate for Apptronik's Apollo 2 humanoid packing auto parts. Rolling out to robotics partners only. Status: vendor blog, partner access, vendor-reported.

What the pattern shows

Three trends stand out across five entries and 28 months.

First, the training data is shifting from robot demonstrations toward human video. RT-2 leaned on web text and images. GR00T N1 mixed robot, human and synthetic data. GR00T 2 puts 100,000 hours of human first-person footage at the center and releases it. If that approach holds, the bottleneck moves from building robots to filming people.

Second, access is diverging. NVIDIA positions its models as open, though GR00T 2 has no license text yet. DeepMind's models stay with partners. That makes outside verification possible for one camp and unlikely for the other.

Third, the evidence standard has not moved. Every headline number is vendor-reported, and the best industrial figure, 96% on a Mercedes-Benz pilot, comes without a published sample size. Simulation benchmarks are improving as shared tests: NVIDIA's partners introduced DexBench and VABAR, and Google released ASIMOV-Jailbreak for safety. Shared tests are the infrastructure a replication culture needs.

Open questions we are watching

• Replication: Will anyone outside NVIDIA or DeepMind publish results on tasks the authors did not choose once GR00T 2 code ships in December?
• Licensing: Will the GR00T 2 license allow commercial use, and will weights be released?
• Production use: Foxconn, Siemens and Nexuni are testing GR00T 2 and none has committed to production, per The Next Web. A signed deployment with failure rates would be a real milestone.
• Safety: DeepMind reports 33% higher safe task completion around obstacles and 43% around nearby people. We want independent measurement of how often a robot follows a disguised unsafe instruction.
• Cost: No lab has published the price of running these models per task or per robot-hour.

For the general-purpose model behind DeepMind's robotics work, see our look at Gemini 4 Argon's robotics claims.

How we update this page

We add an entry when a lab publishes a model card, paper or primary announcement with a dated claim, and we mark it vendor-reported until an outside group replicates it. We correct entries openly when numbers change. If you think we have missed a release, the test is simple: a primary document with a date, and a model that controls a physical robot.