Reinforcement Learning With Vision-Proprioception Model for Robot Planar Pushing.

Area of Science:

• Robotics
• Machine Learning
• Computer Vision

Background:

• Robotic manipulation tasks like object pushing require integrating visual and proprioceptive information.
• Existing methods often struggle with generalization to new objects or rely on high-dimensional raw data.

Purpose of the Study:

• To develop an efficient vision-proprioception model for planar object pushing.
• To improve state representation by fusing visual features and robot state.
• To achieve robust and generalizable object pushing capabilities.

Main Methods:

• Utilized a Variational Autoencoder (VAE) for compact visual feature extraction.
• Fused VAE latent representations with robot end-effector position for state representation.
• Employed Soft Actor-Critic (SAC) algorithm for policy training in simulation.
• Incorporated Hindsight Experience Replay (HER) to enhance sample efficiency.

Main Results:

• The proposed model significantly outperformed a state-based baseline in pushing performance.
• Achieved superior results compared to state-of-the-art policies using raw image observations.
• Demonstrated successful generalization to unseen objects in real-world experiments.

Conclusions:

• The integrated vision-proprioception approach provides an effective state representation for robotic pushing.
• The model exhibits strong generalization capabilities, adapting to novel objects without retraining.
• This method offers a promising direction for robust and versatile robotic manipulation.