Economical Precise Manipulation and Auto Eye-Hand Coordination with Binocular Visual Reinforcement Learning

Yiwen Chen, Sheng Guo, Lei Zhou, Xianyao Ng, Marcelo H. Ang Jr.arXiv 2022 · DOI · arXiv

Abstract

This paper introduces Binocular Alignment Learning (BAL), a reinforcement learning framework designed to enable high-precision robotic manipulation using low-cost hardware. By utilizing two standard webcams and a novel self-calibration technique, the system achieves sub-millimeter precision in tasks such as peg-in-hole insertion and screwing. The approach leverages a minimal simulation environment and domain randomization to bridge the reality gap, resulting in a robust, calibration-agnostic system that outperforms human operators in specific coordination tasks.

Introduction

In industrial robotics, achieving high precision typically necessitates expensive 3D/RGB-D sensors, high-end force/torque feedback, and extensive manual calibration. These setups are not only costly but also fragile; minor physical shifts in camera positioning can lead to system failure. This research addresses these limitations by proposing a bio-inspired binocular vision system that automates eye-hand coordination. By mimicking human binocular vision, the system aims to provide a robust and economical alternative to traditional high-precision sensing.

The binocular hardware setup using two low-cost webcams mounted on a robotic arm.

Methodology

The BAL architecture is built upon three primary components that facilitate autonomous learning and execution:

  1. Self Action Mapping Learning (SAML): This phase automates calibration by observing the relationship between small, known robot movements and the resulting visual changes in the camera frames. The Inverse Mapping Learning (IML) variant optimizes this by using inverse motion vectors to stabilize the neural network's learning process.
  2. Points Alignment Learning (PAL): Manipulation tasks are framed as a "Points Alignment" problem. Using a pre-trained YOLO network, the robot identifies the tool (controllable point) and the target (hole). A Proximal Policy Optimization (PPO) algorithm then learns the necessary movements to align these points.
  3. Minimal Simulation for Points Alignment (MSPA): To accelerate training and ensure robustness, the authors utilized a lightweight simulation environment. Domain randomization of camera parameters during training ensures the policy remains effective regardless of exact physical mounting positions.
The bio-inspired visual design and the workflow from minimal simulation (MSPA) to physical task execution.

Experimental Results

The system was evaluated on a dual-arm KINOVA MOVO robot across several high-precision tasks. The BAL+IML policy demonstrated significant success rates:

  • Insertion: 91.9% ± 1.9%
  • Screwing: 93.7% ± 6.3%
  • Pick-and-Place (Small Items): 100%

Key Insights

  • Binocular Advantage: Binocular IML achieved a 96% success rate in simulation, compared to only 36% for monocular setups, highlighting the necessity of depth perception via two viewpoints.
  • Human vs. Robot: In a comparative test, a human operator using the same visual feed achieved only a 30% success rate, whereas the BAL-trained robot exceeded 90%. This suggests the RL agent is more capable of internalizing complex, non-intuitive visual mappings.
  • Cost Efficiency: The entire visual setup costs approximately $20, a fraction of the $350–$500 required for traditional 3D camera or force-sensing configurations.

Discussion and Conclusion

The BAL framework proves that high-precision manipulation is achievable with minimal hardware costs through intelligent self-calibration and reinforcement learning. While the current system requires human assistance to define the insertion axis, future work will focus on automating line detection and exploring more complex tasks like cloth manipulation. The system's robustness to camera displacement and its sub-millimeter accuracy mark a significant step toward more accessible and resilient industrial automation.

Summary updated

BibTeX
@misc{chen2022economical,
  author = {Yiwen Chen and Sheng Guo and Lei Zhou and Xianyao Ng and Marcelo H. Ang Jr.},
  title = {{Economical Precise Manipulation and Auto Eye-Hand Coordination with Binocular Visual Reinforcement Learning}},
  year = {2022},
  eprint = {2205.05963},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2205.05963},
}
Figure