D-REX: Differentiable Real-to-Sim-to-Real Engine for Learning Dexterous Grasping

TL;DR

This paper introduces a differentiable engine that uses visual observations and control signals to identify object mass and improve robotic grasping policies, effectively bridging the simulation-to-real gap.

Contribution

It presents a novel real-to-sim-to-real engine leveraging Gaussian Splat representations for mass identification and policy learning from limited data.

Findings

Accurately identifies object mass across various geometries.

Enhances grasping policy performance with optimized mass parameters.

Reduces sim-to-real transfer gap effectively.

Abstract

Simulation provides a cost-effective and flexible platform for data generation and policy learning to develop robotic systems. However, bridging the gap between simulation and real-world dynamics remains a significant challenge, especially in physical parameter identification. In this work, we introduce a real-to-sim-to-real engine that leverages the Gaussian Splat representations to build a differentiable engine, enabling object mass identification from real-world visual observations and robot control signals, while enabling grasping policy learning simultaneously. Through optimizing the mass of the manipulated object, our method automatically builds high-fidelity and physically plausible digital twins. Additionally, we propose a novel approach to train force-aware grasping policies from limited data by transferring feasible human demonstrations into simulated robot demonstrations. Through comprehensive experiments, we demonstrate that our engine achieves accurate and robust performance in mass identification across various object geometries and mass values. Those optimized mass values facilitate force-aware policy learning, achieving superior and high performance in object grasping, effectively reducing the sim-to-real gap.