Cybernetic-Labs/openarm-packingbench-v1-groot-n17-90k-350ep

OpenArm PackingBench v1 — GR00T N1.7 (90k, 350-ep)

Fine-tune of NVIDIA GR00T N1.7 (3B) for a bimanual OpenArm pick-and-pack task. This repo holds the inference checkpoint at 90,000 steps.

Base model

• nvidia/GR00T-N1.7-3B (vision-language backbone: nvidia/Cosmos-Reason2-2B)

Dataset

• openarm_dataset_1_groot_lerobot_350_mixed (LeRobot v2.0)
• 350 episodes · 137,678 frames
• 4 cameras: front, high_opposite, wrist_left, wrist_right
• 16-DOF bimanual OpenArm
• 3 task variants (pack one / pack two / close-the-box phrasings)

Training

Setting	Value
Steps	90,000 (this checkpoint)
Global batch size	8
Gradient accumulation	1
Save interval	every 30k (30k / 60k / 90k)
Embodiment tag	new_embodiment
Final train loss	~0.033
GPU	1× RTX 6000 Ada (49 GB)

Launched with gr00t/experiment/launch_finetune.py (Isaac-GR00T n1.7-release, commit 23ace64), absolute modality config, --save-only-model.

Files

• model-0000{1,2,3}-of-00003.safetensors + model.safetensors.index.json — weights
• config.json, experiment_cfg/, processor_config.json, embodiment_id.json — model/inference config
• statistics.json — normalization statistics
• trainer_state.json, training_args.bin — training metadata

Evaluation

Evaluated in Isaac Sim via a GR00T policy server + simulated rollout (two-object packing scene: ShippingBoxSmall + Stapler6). Inference uses --exec-horizon 8 with the OpenArm 4-cam bimanual data config.