About Me
I am a robotics researcher dedicated to achieving true cognitive autonomy—developing systems that do not merely execute commands but possess the capacity for deep understanding, dynamic adaptation, and proactive self-correction. My current work centers on TIWM (Tokenized Intent World Model), a pioneering sparse intent-token interface bridging the gap between raw perception and high-level agency. My long-term goal is to unify interpretable cognitive templates with autoregressive long-horizon planning into a hybrid paradigm for generalizable, human-aligned robotic intelligence.
- Cognitive Robotics
- World Representation
- Deep Reinforcement Learning
- Robotic Middleware (ROS/ROS2)
- LLM/VLM for Robotics
MEng Computer Technology
Central China Normal University
MSc Computer Science
University of Wollongong
BSc Computer Science and Technology
Huaiyin Institute of Technology
- Sparse Semantic Grounding & World Representation
I explore the distillation of minimal, semantically grounded tokens from high-dimensional perceptual streams. This sparse alignment mechanism leverages ‘fuzzy-memory’ to maintain cognitive continuity across heterogeneous environments, bypassing the need for computationally intensive dense reconstructions.
- Cognitive Skill Templates
I design composable action schemas that synthesize symbolic abstraction with Reinforcement Learning (RL) policies. These templates provide a framework for contextual reasoning and rapid adaptation, ensuring robust error recovery and “self-healing” behaviors during complex, non-deterministic human–robot interactions (HRI).
- Intent Autoregression & Spatiotemporal Co-evolution
I leverage intent token autoregression to facilitate the “mental simulation” of future states. This approach enables long-horizon joint spatiotemporal planning by evolving decisions within a sparse latent space, achieving the foresight of dense future rollouts with the efficiency of a tokenized architecture.
