DynaHMRC is a decentralized, role-aware LLM framework that coordinates heterogeneous robot teams through leadership bidding, leader election, and reflective execution, adapting efficiently to dynamic tasks with fewer action and communication steps.

DBPO shifts iterative refinement from inference to training for native one-step generative robot policies, enabling up to 100× faster inference and stable online policy optimization for high-frequency control.

SoPE maps 3D tokens into spherical coordinates and mixes multi-scale frequencies to preserve geometry and direction, improving spatial perception in 3D LVLMs.

C²RoPE improves 3D multimodal reasoning by preserving spatio-temporal continuity in visual tokens and modeling spatial causal relations with Chebyshev causal masking.

CalibWorkflow is a generalizable extrinsic calibration framework that uses MLLMs as visual guides to search, optimize, and refine parameters, enabling robust, hardware-agnostic calibration across diverse sensors and environments.

This paper presents STDArm, a system that transfers static-trained visuomotor policies to dynamic robots, achieving precise and stable manipulation through real-time action correction without retraining.

LLMs are used to realize the collaboration of multiple heterogeneous robots (mobile robot, manipulation robot, and mobile manipulation robot), including three tasks: make sandwich, sort solids, and pack objects.

Model the multi-modal expert policy distribution with multiple scenarios and preferences by diffusion model for robot navigation and collision avoidance.

a novel end-to-end DRL-based method, PathRL, that directly outputs navigation paths without relying on the supervised learning paradigm and is competent for a variety of complex scenarios.

Build and deploy robot decision-making and planning systems on different types of mobile chassis, including differential chassis, Ackerman chassis, logistics vehicles, passenger cars, etc. For more details please refer to my resume.