Pick. Scan. Sort.
Team Stereobot: Thomas Gaviard, Haoran Wang & Gabriel Schwab
Summary: Two arms, one model to rule them both. Our bi-manual robot picks, scans, and sorts packages autonomously—an end-to-end, modular solution built for accurate and scalable warehouse automation.
time-lapse.mp4
Our mission demonstrates a real-world warehouse automation use case through a candy warehouse scenario:
- A bi-manual robotic system autonomously picks and scans items with one arm
- Hands off to a second arm for accurate sorting and packaging
- Showcases a modular, end-to-end solution for high-throughput fulfillment
Our approach is novel in several ways:
| Feature | Description |
|---|---|
| Unified Model | Single model coordinates bi-manual robotic tasks with seamless handoff |
| Modular Design | Independent of specific scanning or sorting technologies |
| High Accuracy | Maintains precision across different objects and scenarios |
| Low-Cost & Adaptable | Same framework deploys across different hardware setups |
teleop.mp4
We trained ACT on a compact dataset using the LeRobot training recipe:
| Parameter | Value |
|---|---|
| Episodes | 150 (+ 30 fine-tuning) |
| Cameras | 4 (top, scan-state, 2× arm-mounted) |
| Training Steps | 35K |
| Hardware | AMD MI300X |
To improve robustness, we fine-tuned the model on 30 failure-case episodes.
| Component | Specification |
|---|---|
| Platform | AMD Ryzen AI 9 HX370 PC |
| OS | Ubuntu 24.04 |
| ROCm | v6.3+ |
| PyTorch | v2.7.x |
| LeRobot | v0.4.1 |
- ✅ Generalization — Successfully generalizes to trained and unseen objects
- ✅ Flexibility — Independent of specific scanning or warehouse setups
- ✅ Simple Control — Requires only color feedback, inference script, and items to pick
High accuracy demonstration of our solution:
demo.mp4
| Resource | Link |
|---|---|
| 🏋️ Model Weights | HuggingFace - amd-act-v3 |
| 📊 Core Dataset | HuggingFace - amd-bimanual-core |
| 📊 Fine-tune Dataset | HuggingFace - amd-bimanual-finetune |