sotata/act-okura-pick-tree-06152026

ACT — okura-pick-tree (Unitree G1)

ACT (Action Chunking Transformer) policy trained with LeRobot for an okra-picking task on a Unitree G1, as part of the Orboh × Toyota Body Research PoC.

Model details

Policy	ACT (policy.type=act)
Dataset	sotata/okura-pick-tree-20260615 (LeRobot v3.0)
Robot	Unitree G1 (Unitree_G1_Dex1)
Observation	2 cameras — observation.images.cam_high, observation.images.cam_right_wrist (480×640×3) + 16-dim observation.state
Action	16-dim (chunk_size 100)
Training	100,000 steps, batch_size 8, final loss ~0.025
Hardware / time	AWS EC2 g5.2xlarge (NVIDIA A10G 24GB), ~4.6 h
W&B run	https://wandb.ai/soutamiyajima4-/lerobot-act-okura/runs/h3lefecq

⚠️ I/O differs from the earlier right-arm model sotata/act-okura-pick-right-06112026 (8-dim, 1 camera). This model is 16-dim with 2 cameras — the inference side (unitree_lerobot) must match this camera/dimension layout.

Usage

Load directly with LeRobot via --policy.path:

python src/lerobot/scripts/lerobot_train.py \
  --policy.path=sotata/act-okura-pick-tree-06152026 \
  ...

or in Python:

from lerobot.policies.act.modeling_act import ACTPolicy
policy = ACTPolicy.from_pretrained("sotata/act-okura-pick-tree-06152026")

The repo contains the full pretrained_model set required for inference: config.json, model.safetensors, and the pre/post-processors (normalizer / unnormalizer) — model.safetensors alone is not sufficient.

Training data

54 episodes / 9,096 frames @ 30 fps. The 100k steps correspond to ~87 epochs over this dataset.

Intended use & limitations

Research artifact for the Orboh × Toyota Body okra-harvesting PoC. Trained on a single task and a small demonstration set; not validated for general deployment. Two-camera, 16-dim G1 setup only.