You can deploy this model and start earning money today!

sotata/
act-okura-pick-06102026

Name: act-okura-pick-06102026
Availability: InStock

sotata/okura-pick-06102026

apache-2.0

Views

No views yet

ACT — okura-pick (Unitree G1, Dex1 grippers)

ACT (Action Chunking with Transformers) policy trained on teleoperated pick demonstrations collected on a Unitree G1 humanoid with Dex1 grippers and a single head camera.

This checkpoint is the last (step 100,000) snapshot in LeRobot pretrained_model format and can be loaded directly with --policy.path=sotata/act-okura-pick-06102026.

Model I/O


Image input	1× RGB `observation.images.cam_left_high` (head camera, 480×640)
State input	16-dim `observation.state` — left arm qpos (7) + right arm qpos (7) + left gripper (1) + right gripper (1)
Action output	16-dim, same layout as state; chunk size 100 (`n_action_steps=100`)

Training


Framework	unitreerobotics/unitree_lerobot (LeRobot 0.4.1 submodule), `lerobot_train.py`
Dataset	sotata/okura-pick-06102026 — 50 episodes / 21,693 frames @ 30 fps (3 failed takes excluded, converted to LeRobot format)
Policy	ACT, default config (batch_size 8, chunk_size 100, lr 1e-5)
Steps	100,000 (~2.7 h on NVIDIA A10G, AWS g5.xlarge)
Final train loss	0.050 (from 7.77 at start)
Logs	W&B run

Usage

bash

python src/lerobot/scripts/lerobot_eval.py \
  --policy.path=sotata/act-okura-pick-06102026 \
  --policy.device=cuda

Important: single-camera config patch

The stock G1_DEX1_CONFIG in unitree_lerobot/utils/constants.py assumes 4 cameras. This policy was trained with one head camera only (color_0 → cam_left_high). For dataset conversion and on-robot inference with unitree_lerobot, trim cameras and camera_to_image_key to:

python

cameras=["cam_left_high"],
camera_to_image_key={"color_0": "cam_left_high"},

Files

model.safetensors — policy weights (207 MB)
config.json, train_config.json — policy / training configuration
policy_preprocessor* / policy_postprocessor* — input normalization / output unnormalization processors