Decoupled Sim-to-Real Transfer via Residual Force-Diffusion Adapters will outperform direct domain randomization in contact-rich tasks.

Motivation

Simulating visual data is increasingly accurate, but simulating high-fidelity contact forces remains notoriously difficult (the 'reality gap'). Trying to learn a unified visual-force policy in simulation often fails when transferred to real hardware due to force discrepancies. A decoupled approach could leverage the structural slow-fast design to isolate domain gaps.

Proposed Method

1. Pre-train the 'Slow' visual component of ImplicitRDP purely in simulation using domain randomization. 2. Freeze the visual backbone and initialize the 'Fast' force component with random weights. 3. Collect a small dataset of real-world demonstrations (few-shot). 4. Train only the 'Fast' force branch and a lightweight 'residual adapter' that fuses the latent embeddings, effectively treating real-world force dynamics as a residual correction to the visual plan.

Expected Contribution

A data-efficient pipeline for deploying contact-rich policies that eliminates the need for complex, high-fidelity force simulation physics.

Required Resources

Robotic manipulator with F/T sensors, physics simulator (Isaac Gym/MuJoCo), small set of real-world demo data.

Source Paper

ImplicitRDP: An End-to-End Visual-Force Diffusion Policy with Structural Slow-Fast Learning

View Paper Details →