Principle:ARISE Initiative Robosuite Dynamics Randomization
Metadata:
- robosuite
- Sim-to-Real Robot Learning from Pixels
- Sim_To_Real_Transfer
- Physics_Simulation
- last_updated: 2026-02-15 12:00 GMT
Overview
Technique for randomizing physical simulation parameters (friction, mass, damping, density) to improve dynamic robustness of learned control policies.
Description
Dynamics randomization varies the physical parameters of the simulation including: global properties (medium density, viscosity), body properties (mass, inertia, position, quaternion), geom properties (friction, contact solver params), and joint properties (stiffness, damping, armature, frictionloss). This addresses the dynamic domain gap where simulated physics do not perfectly match real-world mechanics.
Usage
Use when control policies need to transfer to real robots where exact physical parameters (friction, mass) are uncertain.
Theoretical Basis
Physical simulation parameters are never perfectly known. By training across a distribution of plausible physical parameters, the policy learns adaptive control strategies. Perturbation ratios (relative) are used for mass-like quantities; absolute perturbation sizes for damping-like quantities.