A Data-Driven PINN Framework for Humanoid Walking Using LucidSim, Clio, and LLM-Based Task Orientation

کد مقاله : 1659-ISME2025

نویسندگان

آرین سرداری، علی موسوی * دانشگاه صنعتی شریف

چکیده مقاله

Humanoid robotics is rapidly advancing, propelled by innovations in both hardware design and algorithmic control. This paper presents a comprehensive framework that leverages Physics-Informed Neural Networks (PINNs) to model a humanoid robot's kinematics and dynamics for natural walking patterns. The core innovation is an integrated pipeline that combines open-source human motion capture data, PINN-based system identification, and advanced simulation utilities in LucidSim and Clio. We also propose an LLM-based (Large Language Model) task interpretation component that allows high-level semantic instructions to guide the robot's object orientation and locomotion. Our results demonstrate that the proposed pipeline can learn stable, human-like gait patterns and execute object-targeted tasks in real time. We provide extensive details on the model architecture, data preparation, training procedures, and system integration. Experimental validations in simulation show promising transferability to physical hardware, paving the way for highly adaptive, perceptually aware humanoid robots. Such systems exhibit high sophistications and are looked at as open fields of research in the field of dynamics and control of humanoid robots.

کلیدواژه ها

Complex Control, Navigation, PINN, Robot Perception, Robotics

وضعیت: پذیرفته شده برای ارائه شفاهی