I'm a third-year undergraduate at Guangzhou University, majoring in Network Engineering with a focus on AI systems. My research sits at the intersection of reinforcement learning, embodied AI, and efficient model deployment — closing the gap between RL research and real-world robot and edge hardware.
I'm applying for PhD programs (Fall 2027) in Embodied AI / Reinforcement Learning / Robotics.
🔗 Academic Homepage | 📄 CV | 📧 ldz@e.gzhu.edu.cn
Research Interests:
- Reinforcement learning for embodied agents (off-policy algorithms, sample efficiency)
- Cross-platform RL model deployment (GPU / Jetson / CPU)
- Model compression and quantization for RL policies (INT8 / FP16)
- AI safety and adversarial robustness
Featured Project — rl-deploy-bench
A cross-platform RL model deployment and performance benchmarking toolkit. Export Stable-Baselines3 policies to ONNX/TorchScript, build TensorRT engines with FP16/INT8 quantization, and benchmark latency, throughput, and accuracy across x86 GPU, NVIDIA Jetson, and CPU — all from one config-driven CLI with auto-generated HTML reports.
Research value: Provides a reproducible benchmark for studying the accuracy-latency tradeoff of RL policy quantization across hardware platforms.
| Project | PR | Contribution | Status |
|---|---|---|---|
| scikit-learn/scikit-learn | #34981 | Fixed compute_class_weight coercing string labels (e.g. "1") to int, breaking dict lookup for string class weights |
Open |
| vllm-project/vllm | #57400 | Fixed system_fingerprint: null emitted in non-streaming responses when --fingerprint-mode=none; added exclude_none=True to model_dump |
Open |
| Farama-Foundation/Gymnasium | #1719 | Fixed mj_forward not called after mj_step, causing inconsistent body xpos/qpos in observations (e.g. Reacher-v5) |
Open |
| huggingface/trl | #7264 | Fixed entropy_from_logits returning NaN for zero-probability tokens (0 * -inf = NaN); replaced -inf logps with 0 before multiplication |
Open |
| UoA-CARES/cares_reinforcement_learning | #409 | Algorithm docs (DQN, PPO, overview) rewritten against actual codebase API; 44-algorithm index | Under review |
| OWASP/secure-agent-playbook | #28 | Fixed outdated OWASP LLM Top 10 IDs across 5 files | Open |
| redai-studio/Relax | #294 | Docker-free installation guide (bilingual EN/ZH) | Open |
RL and Learning — PyTorch · Stable-Baselines3 · Gymnasium · MuJoCo Deployment — ONNX · ONNX Runtime · TensorRT · TorchScript · INT8/FP16 quantization Edge and Robotics — NVIDIA Jetson (Xavier/Orin) · ROS · CUDA · cuDNN Systems — Python · C++ · Linux · Docker · Git · CI/CD