PhD student · CASIA & Zhongguancun Academy

I study how embodied agents can learn, imagine, and act reliably. My work connects reinforcement learning, vision-language-action models, world models, and AI safety.

• Institute of Automation, Chinese Academy of Sciences (CASIA) — National Key Laboratory of Cognition and Decision Intelligence for Complex Systems; working with Pin Lyu.
• Zhongguancun Academy — advised by Chen Gao.

_{Background: B.Eng. in Automation, Xi’an Jiaotong University (2024) · Graduate research at CASIA / UCAS since 2024.}

Reinforcement Fine-Tuning of Flow-Matching Policies for Vision-Language-Action Models Online reinforcement learning for flow-matching VLA policies. Paper · Code	A Unified Diagnostic Safety Benchmark for Vision-Language-Action Models Diagnosing physical, instruction, and perception risks in robot policies. Paper · Code
Competitive Multi-Agent Game-Theoretic Tracking with Vision-Language-Action Models Competitive multi-agent reinforcement learning for embodied visual tracking. Paper · ECCV · Proceedings	Spiking world model with multicompartment neurons for model-based reinforcement learning Connecting dendritic computation, spiking networks, and world models. Paper · Code

Full publication list →

_{🟣 Aug 2026 – Present · Physical simulation & foundation-model post-training}

• Develop and debug physics simulation environments.
• Work on reinforcement-learning post-training for foundation models.

_{🔵 Mar 2026 – Jul 2026 · Research Intern}

• Study cross-episode memory and in-context learning for frozen VLA backbones.
• Explore memory-augmented human–robot interaction to improve adaptation and sample efficiency.

_{🟠 Dec 2025 – Feb 2026 · Research Intern · CoMaTrack}

• Research competitive multi-agent tracking and game-theoretic interactions for VLA policies.
• Third author of CoMaTrack, published in the ECCV 2026 proceedings.

_{🟢 Oct 2025 – Dec 2025 · AI Safety Research}

• Research lightweight, transferable inference-time safety alignment for LLMs.
• Contributed to single-neuron self-reflection for LLM safety, balancing generation safety and model utility.

• Reviewer — CoRL 2026.
• Electrical Control Lead — XJTU Intelligent Robotics Team (Sep 2022 – Sep 2023). Led embedded control, motor-drive, power, and communication-system development; worked on sensor fusion and visual tracking across 10+ robot prototypes.
• National Undergraduate Innovation Project (Sep 2021 – Dec 2022). Developed an Intel-based driving-safety assistance system, integrating lane-departure and fatigue detection with embedded control and real-time data processing.

Physical Intelligence research notes — a Chinese-language review of the π-series robot foundation models, from architectures and training signals to deployment.

_{Python · JAX · PyTorch · C++ · ROS · MuJoCo · RoboTwin · LIBERO · Isaac Sim · LaTeX}

lvmingyang2024@ia.ac.cn