Online Parallel MPC-Guided Reinforcement Learning Framework for Legged Locomotion
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Traditional end-to-end reinforcement learning (RL) approaches for legged locomotion exhibit unstable convergence, require labor-intensive hyperparameter tuning, and often yield unnatural gait patterns. To address these limitations, we propose a novel training framework that integrates model predictive control (MPC) with RL to guide policy learning. Our core innovation leverages the inherent physical priors in model-based control (kinematics/dynamics models) to generate high-quality reference trajectories online, which serve as guidance for the RL training process. This hybrid approach reduces dependency on hand-crafted reward functions, enhances sample efficiency, improves convergence process, and improves overall locomotion performance. Notably, it synergistically combines MPC's tunable behavioral specifications with RL's robustness to complex environments. The primary technical challenge lies in running computationally expensive MPC for multiple RL agents. We address this by implementing a GPU-accelerated parallel MPC algorithm within the JAX framework, achieving significant computational speedup. To streamline the training pipeline, we integrate this algorithm with the NVIDIA Isaac Lab training platform. Comprehensive evaluations on a quadruped robot demonstrate the framework's effectiveness in simulated environments, validating its advantages in both training performance and locomotion performance.
Fields of Study & Subjects
Robotic Locomotion and Control
Reinforcement Learning in Robotics
Prosthetics and Rehabilitation Robotics
Publication details
- Journal / source
- International Conference on Electrical Automation and Artificial Intelligence (ICEAAI)
- Pages
- 526–531
- Publication date
- 16 January 2026
- Language
- EN