[IEEE T-ASE’25] The motivation of this work is to address the significant challenges of computationally demanding and safety-critical motion planning for large-scale robotic swarms navigating cluttered environments, which can be applied to missions including search and rescue, warehouse automation, and environmental exploration.

[IEEE TSMC: Systems’25] We propose ReSPIRe, which plans informative search-and-tracking trajectories by estimating belief-space information gain with sigma-point mutual-information approximation and accelerating online belief-tree search through hierarchical particles and reusable rollouts.

[IEEE RAM’25] We propose AINav, an LLM-based adaptive interactive navigation framework that combines primitive skill tree planning, RL-pretrained locomotion and interaction skills, and adaptive replanning to proactively modify cluttered environments and reach originally unreachable goals.

[ICML’24] VoroNav is a zero-shot object navigation (ZSON) framework that builds a Reduced Voronoi Graph from a real-time semantic map to extract traversable paths and intersection-based waypoints. It converts the topological structure into path descriptions and farsight descriptions (via BLIP image captioning), which an LLM (ChatGPT) uses with commonsense reasoning to select navigation waypoints.

[R-AL’24] We propose DOZE, a high-fidelity dataset for open-vocabulary zero-shot object navigation in dynamic environments, featuring 10 realistic 3D scenes and over 18K tasks. DOZE introduces moving humanoid obstacles, diverse open-vocabulary and attribute-varied target objects, scene-text hints, dynamic goal objects, and agent–moving-obstacle collision detection, enabling more realistic evaluation of navigation efficiency, safety, and object recognition in embodied AI.

[IROS’24] We propose SwarmPRM, a hierarchical, highly scalable, computationally efficient, and risk-aware sampling-based motion planning approach for swarm robots. At the macroscopic stage, the collective swarm is represented by a GMM. At the microscopic stage, individual robots track the GMM via a local tracking control law.