清华大学精密仪器系直博生，面向 Robot Learning / Embodied AI Research Engineer 与机器人算法岗位。
I build robot-learning systems from training to trustworthy closed-loop evaluation: multi-GPU VLA training, policy-runtime freshness, metric 3D perception, and causal failure diagnosis.
| Project | What I contributed | Evidence |
|---|---|---|
| Reliable VLA Fine-Tuning & Evaluation | Native SmolVLA/LIBERO contract, 4-GPU DDP, official rollout evaluation, paired flow-noise control | 3.82× DDP speedup; found an RNG confound that changed the apparent scaling verdict |
| Latency-Aware Robot Policy Runtime | End-to-end action provenance and SYNC/FIFO/LATEST scheduling | At 100 ms latency, p95 action age 1.24 s → 150 ms and success 20% → 100% at ~20 Hz |
| Metric 3D Visuomotor Learning | Calibrated RGB-D→world-frame cloud, PointNet control, robustness and multi-view evaluation | 95% success with 50 demos; strong occlusion 3.3% → 100% with two views |
| XR-1 VLA Temporal Alignment | Causal self-prefix plumbing and a 4-GPU real-model counterfactual benchmark | 190 generations on a 5.1B VLA; showed native prefix support alone did not provide latency robustness |
During an industry-practice placement at Goertek, I developed and delivered a multi-person motion-perception and identity-recognition pipeline on a Windows RTX 3050 edge workstation.
- YOLO Pose + OpenCV perception with causal temporal event logic
- FP16, temporal micro-batching, and asynchronous acquisition at approximately 30 FPS
- Matched 141/145 manually labeled events in a three-person historical video
- YuNet/SFace identity binding and UDP/OSC integration with Unity
My Ph.D. work in complex inertial sensing develops transferable strengths in dynamic modeling, numerical simulation, calibration, signal processing, error analysis, and hardware-software system debugging. I apply the same discipline to robot-policy evaluation: isolate contracts, design controls, measure closed-loop behavior, and state failure boundaries explicitly.
Robot learning: PyTorch, LeRobot, SmolVLA, ACT, Behavioral Cloning, Diffusion Policy, action chunking, closed-loop rollout
Robotics and perception: ROS2, MoveIt, ros2_control, MuJoCo, ManiSkill, LIBERO, metric 3D, multi-view fusion, kinematics
Systems: Python, C++, Linux, CUDA, DDP, DeepSpeed ZeRO, Docker, Git
Scope note: the selected robotics projects are simulation or real-model evaluation studies unless a repository explicitly states otherwise; they are not presented as physical-robot deployment claims.