Real2Sim & Sim2Real — Genesis-based Physical Learning

We develop a unified Real2Sim / Sim2Real pipeline built on the Genesis physics engine, combining classical physics calibration, neural-physics correction, and residual learning across both supervised and reinforcement learning regimes. Our long-term goal is a closed-loop framework where (i) real-world rollouts continuously calibrate the simulator, and (ii) policies trained in the calibrated sim transfer back to the real platform with quantifiable robustness.

Recent directions include:

1. State-to-State (ST2ST) and Trajectory-to-State Mappers: Lightweight neural mappers that align Genesis dynamics to observed real-world trajectories, reducing sim-to-real gap before policy learning starts.
2. Residual Learning across Supervised + RL: A behavior-cloned base policy is augmented by a residual reinforcement-learning head that compensates for covariate shift and unmodeled dynamics, achieving sub-millimeter trajectory drift on Genesis solver-optimized environments.
3. Large-scale Multi-Agent Combat Simulation: 3v3 tank-warfare environment with curriculum-based multi-agent reinforcement learning, scaling to thousands of parallel simulated worlds and credit-assignment designs that survive sparse rewards.
4. Neural Physics: Learned correction terms that augment classical simulators where calibration alone is insufficient (terrain, contact, suspension dynamics).

Genesis-based simulation environments (left) and multi-agent tank training (right).

programming experience

Python, PyTorch, Genesis, Isaac Lab, Blender, RL frameworks (PPO/QMIX), behavior cloning + residual RL pipelines