# Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware

**Authors:** Tony Z. Zhao, Vikash Kumar, Sergey Levine, Chelsea Finn  
**arXiv:** [2304.13705](https://arxiv.org/abs/2304.13705)  
**Date:** 2023-04-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.13705v1/

## Key points

- Fine manipulation tasks in robotics require precision, coordination of contact forces and visual feedback.
- These tasks typically require high-end robots, accurate sensors or careful calibration which makes them expensive and difficult to set up.
- Recent developments in learning have enabled low-cost and imprecise hardware to perform these tasks.
- The paper titled "Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware" presents a low-cost system that performs end-to-end imitation learning directly from real demonstrations collected using a custom teleoperation interface.
- The authors develop an algorithm called Action Chunking with Transformers (ACT), which learns a generative model over action sequences to address challenges in high-precision domains.
- ACT allows the robot to learn six difficult tasks in the real world with 80-90% success rates after only 10 minutes worth of demonstrations.
- Tasks include opening a translucent condiment cup and slotting a battery.
- This research shows promise for enabling low-cost robots to perform fine manipulation tasks previously reserved for high-end machines.

## Summary

The field of robotics has long struggled with fine manipulation tasks, such as threading cable ties or slotting a battery, which require precision, careful coordination of contact forces, and closed-loop visual feedback. Typically, performing these tasks requires high-end robots, accurate sensors, or careful calibration, making them expensive and difficult to set up. However, recent developments in learning have enabled low-cost and imprecise hardware to perform these tasks. In their paper titled "Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware," Tony Z. Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn present a low-cost system that performs end-to-end imitation learning directly from real demonstrations collected using a custom teleoperation interface. The authors acknowledge that imitation learning presents its own challenges in high-precision domains: errors in the policy can compound over time and human demonstrations can be non-stationary. To address these challenges, the authors develop a simple yet novel algorithm called Action Chunking with Transformers (ACT), which learns a generative model over action sequences. ACT allows the robot to learn six difficult tasks in the real world with 80-90% success rates after only 10 minutes worth of demonstrations. These tasks include opening a translucent condiment cup and slotting a battery. Overall, this research shows promise for enabling low-cost robots to perform fine manipulation tasks previously reserved for high-end machines. The project website provides further details on the methodology used by the authors and their results demonstrating the potential of this approach to enable low cost robots to carry out complex manipulation tasks.