squaredcuber/asimov-groot-n17

GR00T N1.7 for the Asimov 1 humanoid (simulation)

A fine-tune of NVIDIA GR00T N1.7-3B for Menlo Research's Asimov 1, an open-source humanoid, in the configuration that actually ships: 25 DOF, passive printed "dummy hands" (no grippers), one monocular head camera. As far as we know this is the first GR00T policy for Asimov.

Simulation only. It has not been run on a real robot.

Tasks and results

Three language-conditioned tabletop tasks in MuJoCo. Evaluation is closed loop on 50 held-out layouts per task, with seeds disjoint from every training seed. The policy runs at 20 Hz, predicts 16-step chunks and executes 8 of them per call. Success must hold for 0.5 s.

Task	Instruction example	ckpt 6000	ckpt 10000	95% CI (10000)
Push (one fist)	"push the purple block onto the orange pad"	46%	84%	0.71–0.92
Button	"press the red button"	36%	64%	0.50–0.76
Box (bimanual squeeze, carry, place)	"lift the box and put it on the shelf"	2%	56%	0.42–0.69

• Main failures at step 10000:
  • push: block short of the pad (5), block never moved (3)
  • button: no press (10), partial press (6), wrong button (2)
  • box: dropped or not placed (16)
• Not converged: success was still rising steeply between steps 6000 and 10000, so longer training should help.
• Open-loop on training episodes: action MAE 0.027 rad.
• Inference: 95 ms per call on an L40 (PyTorch eager).

Example rollouts (one success and one failure per task) are in results/closedloop_checkpoint-10000/videos/.

Training

• Data: 900 scripted-expert demonstrations, 300 per task. That is 155k frames at 20 Hz, 43 instructions, randomized object poses, colors, distractors, lighting and camera jitter.
• Observation:
  • head camera, 256×256
  • joint positions: left and right arm (5 each), waist yaw, neck yaw and pitch
  • the instruction
• Action: absolute joint targets for the same 13 DOF, learned as RELATIVE (delta from current state), 16-step horizon.
• Fine-tune: NEW_EMBODIMENT, default GR00T recipe (projector and action head trained, backbone frozen). 10,000 steps at batch 64, learning rate 1e-4, color jitter. One L40 for 2 h 43 m. Loss went from 0.095 to 0.010.
• Checkpoints: checkpoints/checkpoint-{2000..10000}, with weights and processor but no optimizer state.

Simplifications and assumptions

These are deliberate scope cuts for v1:

• Pelvis welded: legs held at the standing pose. Balance is the locomotion controller's job on hardware.
• Hand collision shapes: convex hulls of Menlo's printed dummy-hand files. Mass (0.18 kg) and mounting are assumptions.
• Head camera: mounted pitched 45° down inside the head with a 75° vertical FOV. The neck's 19.5° downward limit cannot see the table on its own. This needs checking against the real camera mount.
• Box task: the carton is held by friction only (μ 0.5, no welds). It uses a 0.30–0.34 m carton because Asimov's fists only face each other in a narrow band, set by one-sided shoulder roll and a 5-DOF arm.

Code

Simulation tasks, scripted experts, dataset export, GR00T modality config, closed-loop evaluation and the one-shot training script: asimov-groot (to be published).

License

Licensed by NVIDIA Corporation under the NVIDIA Open Model License. This is a derivative of nvidia/GR00T-N1.7-3B, which uses nvidia/Cosmos-Reason2-2B, under the same license. The Asimov robot model is from menloresearch/asimov-1 (GPL-2.0 software / CERN-OHL-S-2.0 hardware).