Principle:ARISE Initiative Robosuite Camera Randomization
Metadata:
- robosuite
- Domain Randomization for Sim2Real
- Sim_To_Real_Transfer
- Computer_Vision
- last_updated: 2026-02-15 12:00 GMT
Overview
Technique for perturbing camera parameters (position, rotation, field of view) to improve viewpoint robustness of learned policies.
Description
Camera randomization introduces small perturbations to camera position, orientation, and field of view. This forces the policy to be robust to slight camera placement errors and calibration differences between simulation and real-world setups. Perturbation sizes are additive to the default camera parameters.
Usage
Use when camera placement in the real world may differ slightly from simulation, or when using eye-in-hand cameras that have variable positions.
Theoretical Basis
Camera viewpoint variation affects how objects appear in images. By training across a distribution of camera parameters, the policy learns viewpoint-invariant features. Position perturbation size should match expected real-world calibration error. Rotation perturbation in axis-angle representation.