The content discusses the challenges of representing and controlling contact-rich robotic systems, proposing a method to reduce hybrid model complexity while maintaining high performance. The approach is demonstrated on synthetic systems and a three-fingered robotic hand manipulating an unknown object.
The authors focus on learning reduced-order models to achieve real-time control in dexterous manipulation tasks. They introduce a trust-region LCS model predictive controller and iterate through training the reduced-order LCS and updating the rollout buffer. The algorithm aims to minimize the task performance gap between full-order dynamics and reduced-order models.
The theoretical analysis justifies the learning process by showing that fitting a reduced-order model well to full-order dynamics can lead to similar task performance. The practical algorithm involves training the reduced-order LCS, setting trust regions, and running trust-region LCS MPC policies on the robot system.
Key points include identifying fewer task-relevant hybrid modes, using model predictive control for real-time control, and demonstrating state-of-the-art closed-loop performance in dexterous manipulation tasks.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Wanxin Jin,M... at arxiv.org 02-29-2024
https://arxiv.org/pdf/2211.16657.pdfDeeper Inquiries