I build simulation pipelines for robot learning and manipulation, with open-source work on numerical correctness and GPU memory efficiency.
Engineering Notes · GitHub · Contribution Ledger
10 merged upstream PRs across Newton, MuJoCo Warp, PyPose, and COLMAP.
Selected PR evidence
- Newton: IK correctness #4871 · Autodiff L-BFGS #4874 · Cloth assembly #4942
- MuJoCo Warp: Contact reductions #1853
- PyPose: Broadcast views #439
- COLMAP: BA controls #4892 · Driver diagnostics #4903
View All Merged PRs · Verified contribution ledger
Upstream merges only above; personal repositories and fork merges are listed separately in the ledger.
Project descriptions and metrics below are self-reported and have not been independently verified against public code or experiment records. Illustrations are method schematics.
Reconstruct egocentric dexterous manipulation and retarget hand–object motion using DaID, MPPI physical search, and PPO feedback control.
| Reported result | Evaluation scope |
|---|---|
| 4.41× faster · 1,564 → 354 s | Reference construction for 276 frames |
| 34.4% → 73.4% basket success | Fixed development set |
Recover scene geometry and materials with RGB-D, 3DGS, meshes, differentiable rendering, and PBR.
Engineering scope: metric dimensions, collision-proxy repair, and Lambert + GGX material optimization.
Build planning and physical demonstration pipelines with SIMPLE, CuRobo, and G1 + O6/O7; fine-tune GR00T and evaluate closed-loop behavior.
| Reported result | Evaluation scope |
|---|---|
| 1,000 successful trajectories | Physical manipulation demonstrations |
| 86.7% simulation / 60% real-robot success | No real-robot training data |
- Robotics: Robot kinematics · collision detection · MPPI · CuRobo
- Physics: MuJoCo · Newton · NVIDIA Warp
- Learning: PPO · GR00T · VLA · Sim2Real
- Graphics: RGB-D · 3DGS · differentiable rendering · PBR
- Engineering: Python · C++ · Git · regression testing



