Nvidia has open-sourced Flora, the system it uses to label robot training video for its GR00T humanoid-robot models. The suite, published by Nvidia's GEAR robotics lab with a research paper dated Tuesday, Oct. 20, 2026, is three models that watch robot and human manipulation video and write the labels robot AI needs: when each action starts and ends, what the action was, and whether it went well. The models are free on Hugging Face for non-commercial research.

What's in the release

  • Flora-TempSeg splits a video into action intervals, marking where each step begins and ends.
  • Flora-Captioner (27 billion parameters, post-trained from Alibaba's Qwen3.5-27B) writes one short instruction for each interval, such as "Place the blue mug on the tray with your right hand," according to its Hugging Face model card.
  • Flora-Rewarder (28 billion parameters) scores clips for outcome and process rewards, including against rubrics a user writes, the signals used in reinforcement-learning post-training and for monitoring robots in the field.

The lab says Flora is "battle tested at scale, labeling 200,000 hours of data per day across 512 GB300s," Nvidia's Blackwell Ultra chips, and that it was the data engine used to train GR00T-2. The paper's authors include Nvidia robotics leads Linxi "Jim" Fan and Yuke Zhu.

How it compares

Nvidia says Flora posts top scores on the WGO-Bench captioning benchmark and the RoboReward and Robometer reward benchmarks, is competitive with OpenAI's GPT-6 Astra on its internal tests, and runs roughly 125 times cheaper than that API. The paper is candid about trade-offs: on its dense-reward test Flora-Rewarder caught 53.3% of failure steps, where no other model topped 11.7%, but it also wrongly flagged about 11% of normal steps as failures. On a rubric-scoring test, GPT-6 Astra was more accurate (an error score of 3.6 versus Flora's 6.5).

The fine print

Flora is released under Nvidia's non-commercial license, so companies can't put it straight into products. Nvidia's model card also says Flora's outputs are annotations and that it "is not intended to directly control robots or to serve as the sole safety monitor for a physical system."

Why it matters for the humanoid race

Labeled video is one of the slowest, costliest parts of training general-purpose robots, and every humanoid program, from Tesla's Optimus to Figure and OpenAI's planned robot, is chasing more of it. By giving away its labeling pipeline, Nvidia makes it cheaper for research labs and rivals to build training sets on its tools. For more on the field, see OpenAI's plan to demo a humanoid in 2027 and Nvidia's physical-AI partnership with Germany, ahead of Jensen Huang's GTC Berlin keynote on Oct. 21.