industrial-claw-groot-specialists

Industrial Claw GR00T N1.7 DROID specialists

Public inference weights for the Industrial Claw two-Franka Isaac Sim pick-and-place demonstration. These checkpoints were post-trained from nvidia/GR00T-N1.7-DROID at immutable revision 05e7cc97e40dbd33b0890c35cc0214fcb0547ab5.

Checkpoints

Each directory is a complete PyTorch/BF16 GR00T inference checkpoint with eight required files. The embodiment is OXE_DROID_RELATIVE_EEF_RELATIVE_JOINT.

Training

The three normal specialists each used 168 simulated physical-friction episodes: 84 left-arm and 84 right-arm demonstrations, with the other two objects visible as distractors. The corrective green checkpoint extends the green training set with 32 layout-3 correction episodes (200 total).

Training used NVIDIA Isaac-GR00T source revision e5749287857afd97b78f1147166137de29746392, 1,000 optimizer steps, global batch 16, learning rate 1e-4, cosine scheduling, 5% warmup, weight decay 1e-5, state dropout 0.2, image augmentation, and seed 42. Language and visual backbones remained frozen; the projector, action model, and VLA adapter were trained. The original simulation dataset is not included.

Download and serve

Nested checkpoint folders should be downloaded locally before serving:

hf download mainakmallick/industrial-claw-groot-specialists \
  --include 'red-checkpoint-1000/*' \
  --local-dir ./industrial-claw-groot-models

python gr00t/eval/run_gr00t_server.py \
  --model-path ./industrial-claw-groot-models/red-checkpoint-1000 \
  --embodiment-tag OXE_DROID_RELATIVE_EEF_RELATIVE_JOINT \
  --host 127.0.0.1 --port 5581

Verify downloaded files with SHA256SUMS before inference.

Evaluation scope

The packaged deployment previously validated each normal specialist across six fixed arm/layout cases in Isaac Sim. A fresh 10-rollout red evaluation on 2026-08-03 covered three left-arm positions spanning 0.238 m and achieved 9/10 physically verified pick-and-place results; the remaining rollout was safely rejected because a predicted raw waypoint exceeded a Franka joint limit.

These results depend on the Industrial Claw executor and fixed camera, state, and scene contracts. Layouts 3 and 5 use deterministic joint-space alignment to a recorded pre-grasp pose before GR00T performs the final approach, grasp, transport, and release. This alignment is not RMPflow, but those cases are not pure model control from the home pose. Broad random-spawn generalization and physical-hardware safety are not claimed.

Safety and license

These are research and simulation artifacts, not a safety system. Do not use them to control physical equipment without independent validation, motion limits, collision protection, emergency stopping, and human oversight.

The upstream model weights and these derivatives are governed by the included LICENSE (NVIDIA Open Model License). Retain the license and upstream notices when redistributing. The license file, rather than this summary, controls use.

Upstream

• NVIDIA Isaac-GR00T
• GR00T N1 paper