Context
Before deploying RL or planning algorithms to hardware, we needed safe validation. I built a MuJoCo-based simulation stack that mirrors real robot interfaces. It also served as our visualization and sim-to-real validation layer (same APIs, same controllers, same rollouts).
My Contribution
- Designed simulation architecture matching hardware APIs (non-ROS)
- Integrated controllers and planners identical to the real stack
- Built automated rollout validation pipelines
- Enabled RL/IL policy testing before deployment
Technical details
Key design decision: sim and real shared the same interface layer.
This minimized sim-to-real friction and avoided:
- Interface mismatches
- Timing inconsistencies
- Controller discrepancies
Validated in the stack:
- MPC planners
- RL/IL policies
- Teleoperation pipelines
Impact
- Eliminated robot/environment damage during testing
- Enabled faster iteration on RL and planning algorithms
- Reduced deployment risk significantly
- Improved debugging via reproducible visualization and sim-to-real checks