SO-101 · GR00T N1.7 · MuJoCo twin

Preflight

Know whether the robot will succeed before it moves. One photo of the table becomes a physics twin; the robot's own policy rehearses there dozens of times; Preflight signs off GO RETRY or ASK_HUMAN with the evidence, explained by Nemotron on Nebius Token Factory.

01 · Live check

Check a photo of the mat

Upload a top-down photo of the printed A3 mat (four corner markers, a blue block and a red block). Live, the demo runs perception and the trust gate; rehearsal needs a GPU, so it runs for the scenes below. Uploads are limited per hour.

No photo chosen
Pick and place on
No mat? Try an example environment and instruction:
02 · Rehearsed scenes

Every verdict, with its rollout

Synthetic frames laid out to show each verdict, and one real phone photo of objects the twin cannot place, which the gate refuses to rehearse. Refused scenes are filed to Autopilot.

03 · Benchmark

Across environments, objects and tasks

96 rehearsals per suite (12 scenes × 8 variants) in the randomised twin. The weak rows are the honest ones: new objects are harder, and tasks it was never trained on fail almost every time.

04 · Architecture

What runs where

One photo in, one verdict out. Models are linked from the collection.

StageWhat happensWhere it runs
01Phone photoTop-down frame of the A3 mat with four ArUco corner markers.your browser, uploaded to this Space
02Markers and twinOpenCV fixes the mat in millimetres; objects become a MuJoCo twin of the SO-101 cell.this Space (Hugging Face, CPU)
03VisionA vision-language model names each object; disagreement with the colour detector lowers trust.Nebius Token Factory (Qwen3.8 VL)
04Trust gateLow twin fidelity: ASK_HUMAN before anything is rehearsed, because a rehearsal in a twin that does not match the table measures nothing. A task outside training is flagged and still rehearsed - the success rate is the answer to whether it works.this Space
05RehearsalThe robot's own policy, GR00T N1.7 fine-tuned by DAgger, runs randomised rollouts in the twin.GPU job (job.Dockerfile); A100 for the gallery
06VerdictGO / RETRY / ASK_HUMAN from the success rate and its 95% interval, in plain words.Nemotron 3 Ultra, Nebius Token Factory
→AutopilotWhat fails here becomes the next training cycle; a paired gate decides, a human approves.Autopilot Space; A100 trains