g1-inspire-redball-n16

g1-inspire-redball-n16

GR00T N1.6 fine-tune for the red-ball pick-and-place task on a Unitree G1 with Inspire hands. Trained on 76 teleop episodes (IsaacLab, 424×240 ego_view, 34-frame chunks at 60 Hz).

This is the N1.6 counterpart of birbirll/g1-inspire-redball-n17, trained with the same recipe as birbirll/g1-inspire-piston-n16: effective batch 32 (8×4 accumulation), 10k steps, lr 1e-4 cosine, state_dropout 0.8, color jitter, paged_adamw_8bit + gradient checkpointing (fits a 24 GB GPU). Final train loss 0.016.

Serving

Serve this repository root directly — processor_config.json and both model shards live at the root (do not point at a checkpoint-N subdir):

python gr00t/eval/run_gr00t_server.py \
    --model_path <path-to-this-repo> \
    --embodiment_tag NEW_EMBODIMENT --port 5555

Serving rules that matter:

• Embodiment tag is NEW_EMBODIMENT (not UNITREE_G1).
• Low-rate serving: run the client at inference_hz 1.25 (chunked, no temporal ensembling). N1.6 checkpoints from this recipe under-commit when replanned faster with ensembling.
• Camera input: 424×240 ego_view, raw passthrough — the server resizes via its own processor config (shortest_image_edge=256).
• Action layout (30-D): 14 arm joints, 12 Inspire hand joints, base_height, 3-D navigate command. State 63-D (arms, hands, waist, tactile — tactile unused by the model).

Status

• Open-loop replay against the training data: verified.
• Closed-loop in MuJoCo: carries without grasp (visual domain differs from the IsaacLab training render); closed-loop in the training render: grasp triggers.
• Real-robot deployment: wiring validated on a hardware-free bench; behavior on hardware not yet evaluated.