Battery Sorting Robot — ACT Policy
Trained ACT (Action Chunking with Transformers) policy for bimanual battery sorting with SO100 robot arms.
The robot picks a target battery from a pile of objects and places it into a basket, achieving 86% average success rate across 200 evaluation trials.
GitHub: alanyukeroo/bimanual-battery-picker
Course: TECHIN 517 — University of Washington, Spring 2026
Team: Alan Nur · Fan Zhang · Wei Chang
The Task
| Step | Description |
|---|---|
| 1. Pull Basket | Left arm grabs the basket and pulls it closer |
| 2. Pick Battery | Right arm grasps the target battery from a pile |
| 3. Place & Reset | Right arm places battery into basket, arms return home |
Evaluation Results
200 trials across 4 test conditions:
| Condition | Success Rate |
|---|---|
| Single Battery (clean workspace) | 90% |
| Distractors — In Dataset | 90% |
| Distractors — Out of Dataset | 80% |
| Similar Batteries | 84% |
| Overall | 86% |
Training Details
| Parameter | Value |
|---|---|
| Policy | ACT (Action Chunking with Transformers) |
| Backbone | ResNet18 |
| Framework | LeRobot |
| Demonstrations | 351 episodes |
| Training steps | 100,000 |
| Training time | ~3 hours on RTX 5090 |
| Action chunk size | 100 |
| Cameras | Overhead RealSense + Left/Right wrist RGB |
| Image resolution | 640×480 |
Usage
# Download the model
hf download alannur/bimanual-battery-picker-act \
--local-dir outputs/train/act-bimanual-battery-v4/checkpoints/last/pretrained_model
# Run on real robot (see GitHub repo for full setup)
bash scripts/eval.sh
See the GitHub repo for full setup and usage instructions.
Hardware
- 2× SO100 follower arms (bimanual)
- 2× USB webcams (wrist cameras)
- 1× Intel RealSense D435 (overhead)
- Downloads last month
- 3