ACT — SO101 Multi-Object Packing (Reproduce)
A reproduction of ACT (Action Chunking with Transformers) (Zhao et al., 2023) trained with LeRobot on the SO101 robot arm multi-object packing task. ~51.7M parameters.
Part of a study reproducing published robot-learning policies to understand and improve on them.
Model details
- Policy: ACT (
act), 51.7M parameters, safetensors - Observation: joint state + 3 cameras (
top,front,right) - Action: action chunk of 100 steps (
n_action_steps=100) - License: Apache-2.0
Training
| Setting | Value |
|---|---|
| Dataset | Qiu-Xinchuan/so101_Multi-object-packing |
| Steps | 100,000 |
| Batch size | 8 |
| Learning rate | 1e-5 |
| Seed | 1000 |
Reproduce with LeRobot:
lerobot-train \
--policy.type=act \
--dataset.repo_id=Qiu-Xinchuan/so101_Multi-object-packing \
--steps=100000 \
--batch_size=8 \
--policy.optimizer_lr=1e-5 \
--seed=1000 \
--job_name=act_packing
Usage
from lerobot.policies.factory import make_policy
from lerobot.configs.policies import PreTrainedConfig
config = PreTrainedConfig.from_pretrained("Porameht/act_so101_packing_reproduce")
policy = make_policy(config, ds_meta=...) # dataset meta from Qiu-Xinchuan/so101_Multi-object-packing
Or evaluate / run on a real SO101 via LeRobot's eval and robot control scripts:
lerobot-eval --policy.path=Porameht/act_so101_packing_reproduce ...
Evaluation
Not yet evaluated on hardware — success-rate numbers will be added after SO101 rollouts. See the companion SmolVLA reproduction: Porameht/smolvla_so101_packing_reproduce.
Citation
@inproceedings{zhao2023learning,
title={Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware},
author={Zhao, Tony Z and Kumar, Vikash and Levine, Sergey and Finn, Chelsea},
booktitle={Robotics: Science and Systems (RSS)},
year={2023}
}