The question
Which policy behaviors are robust, and which depend on assumptions hidden in the simulator or runtime?
The approach
Use IsaacLab with an independent MuJoCo validation path. Test domain randomization and perturbations, then maintain observation and action parity from training through exported ONNX inference to the control loop.
Validation focus
- Vary contact, friction, noise, inertia, latency, and actuator response.
- Check observation preprocessing, action scaling, and control frequency.
- Trace commanded actions against measured responses to distinguish model and integration errors.
Current boundary
Simulator parity and deployment consistency are research objectives under evaluation. Cross-simulator checks are not a substitute for physical validation.