research

LLM+Robotics

The application of Large Language Models (LLMs) in the field of robotics lies at its core in endowing traditional robotic control systems with the general knowledge, commonsense reasoning, and logical planning capabilities inherent in LLMs. Its primary objective is to achieve a paradigm shift鈥攆rom the rule-driven "perception-decision-control" paradigm to a "cognition-understanding-reasoning" paradigm that more closely approximates human thinking.
Taking the autonomous vehicle system as an example, specifically, LLMs, in the behavioral decision-making phase, can understand the intentions of various traffic participants in edge cases and generate human-like interactive behaviors. In the low-level control phase, when combined with traditional control models, LLMs can translate high-level decisions and even human instructions into control commands that exhibit greater environmental adaptability while ensuring safety.

LLM in Robotic Decision-Making LLM in Low-Level Control

Advanced Intelligent Autonomous Flight


While most current works of autonomous flight for micro aerial vehicles (MAVs) is primarily based on a layered 'perception-localization-planning-control' framework, we are more interested in endowing MAVs with human-like cognitive intelligence, rather than modular input-output processing. Our research spans Neuromorphic Learning-based Autonomous Navigation, Decision & Control leveraging Vision-Language-Model (VLM), and Construction & Understanding of Semantic Scene Maps, all of which have been widely deployed and validated on various platforms.

Morphological Intelligent Robot


Morphological Robot LLM in Low-Level Control
Inspired by natural structures such as vertebral columns and cellular architectures, this project focuses on the design and intelligent control of tensegrity-based cross-domain robots to enhance their adaptability and task performance in complex, unstructured environments. The project advances three core technologies: self-adjusting and adaptive integrated mechanism design, intelligent motion control combining analytical mechanics with behavior generation, and task鈥搑obot鈥揺nvironment co-adaptive autonomous decision-making and path planning. The team has developed a prototype that synergistically integrates the stability of rigid structures with the deformability of soft materials, successfully demonstrating multimodal locomotion and exhibiting excellent cross-domain mobility and impact resistance鈥攐ffering an efficient and safe robotic alternative for high-risk exploration missions.