---
library_name: lerobot
license: apache-2.0
pipeline_tag: robotics
tags:
- robotics
- lerobot
- act
- so101
datasets:
- aakashv100/so101-pick-cube-v2
---

# ACT — SO-101 pick cube v2

ACT (Action Chunking Transformer) policy for the SO-101 follower arm.
Task: **"Pick up the cube and place it in the bowl"**.

## Training

- Dataset: [aakashv100/so101-pick-cube-v2](https://huggingface.co/datasets/aakashv100/so101-pick-cube-v2) — 50 teleoperated episodes, 30 fps
- Cameras: `gripper_cam` (wrist, 640x480) + `top_cam` (overhead wide-angle USB, 640x480)
- State/action: 6-DoF joint positions (degrees) + gripper
- Steps: 60k, chunk_size = n_action_steps = 100
- Checkpoint: 060000

## Usage (lerobot)

```bash
lerobot-record \
  --robot.type=so101_follower \
  --policy.path=aakashv100/act_so101_pick_cube_v2 \
  --dataset.single_task="Pick up the cube and place it in the bowl" \
  ...
```

## Evaluation

Evaluated with a 50-trial fixed-position + 50-trial random-position protocol
(same checkpoint, cube start positions within a 10x10 cm zone). Note: all
training episodes start the cube at a single fixed position, so random-position
performance measures out-of-distribution generalization.