ProxIndustries/openarm-gr00t-right-arm-pick-v1

OpenARM GR00T N1.7: right-arm pick (v1)

Fine-tuned from nvidia/GR00T-N1.7-3B on ProxIndustries/aws-openarm-right-arm-pick-slow3-215. That dataset has 215 episodes and 61,249 frames at 30 fps: 200 task episodes and 15 idle episodes, in which the robot stays at the start pose with no bottle on the table.

Task: "Pick up the bottle on the right and place it into the right bowl."

This repo contains checkpoint-15000, inference files only. The optimizer, scheduler and RNG state are not included.

Embodiment

• Embodiment tag: NEW_EMBODIMENT. The modality config is openarm_config.py.
• Cameras: chest_view, left_wrist_view, right_wrist_view (480x640).
• State and action (16-D): left_arm(7) + left_hand(1) + right_arm(7) + right_hand(1).
• Action chunk: 16 steps (~0.53 s). Arms use relative joint actions; grippers are absolute.

Training

• Code: Isaac-GR00T main @ 51d4c89, using examples/finetune.sh.
• Schedule: 15,000 steps at global batch 32 on 1x RTX 6000 Ada, in about 5.7 h (~8 passes over the data).
• Optimizer: AdamW at lr 1e-4 with cosine decay, 5% warmup and weight decay 1e-5.
• Tuned parts: projector, diffusion model and vlln. The LLM and visual encoder are frozen.
• Regularization: state dropout 0.2 and color jitter (brightness 0.3, contrast 0.4, saturation 0.5, hue 0.08).
• Final training loss: ~0.017.

Open-loop eval (training episodes, 4 denoising steps)

Episode	MSE	MAE (rad)
Task, ep 0	9.6e-5	0.0051
Idle, ep 207	4.1e-6	0.0013

On the idle episode, the predicted arm motion per 16-step chunk was at most 0.010 rad. The ground truth is 0. Both episodes are in the training set, so these numbers show fit, not generalization. The model has not yet been evaluated in closed loop.

Usage

hf download ProxIndustries/openarm-gr00t-right-arm-pick-v1 --local-dir <dir>

from gr00t.data.embodiment_tags import EmbodimentTag
from gr00t.policy.gr00t_policy import Gr00tPolicy

policy = Gr00tPolicy(embodiment_tag=EmbodimentTag.NEW_EMBODIMENT, model_path="<dir>", device="cuda")