DROP: Dexterous Reorientation via Online Planning

2024-09-22

arXiv.org

AI Summary

This work addresses the challenges of online planning difficulty and poor sim-to-real transfer in contact-intensive dexterous manipulation. We propose a lightweight, real-time sampling-and-optimization online controller, and—crucially—integrate it end-to-end with RGB-D vision-based pose estimation for the first time, enabling contact-rich manipulation planning without large-scale simulation pretraining. Our method leverages parallel physics simulation and contact-aware motion planning, and is validated on a physical dexterous hand performing cube reorientation. It achieves >95% success rate, matching state-of-the-art offline reinforcement learning methods in performance. The fully deployed pipeline incurs <100 ms end-to-end latency, significantly improving responsiveness and environmental adaptability. This establishes a new paradigm for real-time dexterous manipulation in highly dynamic, strongly interactive scenarios.

Technology Category

• Intelligent Robots: Manipulation
• Planning, Routing, and Scheduling: Replanning and Plan Repair
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web
• User Modeling, Personalization and Recommendation: User modeling and simulation for interactive and conversational systems

Abstract

Achieving human-like dexterity is a longstanding challenge in robotics, in part due to the complexity of planning and control for contact-rich systems. In reinforcement learning (RL), one popular approach has been to use massively-parallelized, domain-randomized simulations to learn a policy offline over a vast array of contact conditions, allowing robust sim-to-real transfer. Inspired by recent advances in real-time parallel simulation, this work considers instead the viability of online planning methods for contact-rich manipulation by studying the well-known in-hand cube reorientation task. We propose a simple architecture that employs a sampling-based predictive controller and vision-based pose estimator to search for contact-rich control actions online. We conduct thorough experiments to assess the real-world performance of our method, architectural design choices, and key factors for robustness, demonstrating that our simple sampling-based approach achieves performance comparable to prior RL-based works. Supplemental material: https://caltech-amber.github.io/drop.

Problem

• Achieving human-like dexterity in robotics
• Online planning for contact-rich manipulation tasks
• Robust sim-to-real transfer in reinforcement learning

Innovation

• Online planning for contact-rich manipulation
• Sampling-based predictive controller
• Vision-based pose estimator