Synergistic Control for Quadrupedal Locomotion: An RL-Augmented MPC Framework for Dynamic Load Adaptation on Unstructured Terrains
Year: 2025
Type: preprint
Abstract: The deployment of quadrupedal robots in realworld scenarios necessitates the ability to perform dynamic locomotion while carrying unknown, shifting payloads over complex, unstructured terrain. This dual challenge remains a significant barrier, as existing control paradigms exhibit fundamental limitations. Model Predictive Control (MPC), while effective for predictive planning with explicit constraints, falters when faced with significant unmodeled dynamics introduced by payloads and environmental interactions. Conversely, Reinforcement Learning (RL) can generate remarkably robust policies for specific environments but lacks formal guarantees and struggles with precision and generalization to novel dynamic tasks. This paper introduces a novel, synergistic control framework that addresses these limitations by integrating a convex MPC planner with an adaptive RL policy. The proposed RL-augmented MPC architecture leverages the strengths of both methodologies: MPC provides a computationally efficient, constraint-aware motion plan based on a simplified rigid-body model, while the RL policy, trained using a privileged learning scheme, operates at a high frequency to provide real-time adaptive corrections. This policy learns to infer the complex, unmodeled dynamics arising from both terrain irregularities and payload interactions purely from proprioceptive sensor data, modulating the MPC's behavior to maintain stability. The efficacy of this framework is validated through extensive, synthesized experiments on a highfidelity model of the Unitree Go1 quadruped in a physics-based simulation. The results demonstrate that the proposed system significantly outperforms standalone MPC and pure RL baselines in metrics of success rate, velocity tracking, energy efficiency, and stability, particularly in its unique ability to stabilize dynamic loads up to 50% of the robot's body mass while traversing rough terrain. This work represents a critical step toward the development of truly versatile, load-bearing legged robots capable of robust operation in unpredictable, realworld environments.
Authors: Chandan M, Chiranth M U
Institutions: Jain University, Dayananda Sagar University
Language: English
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4414200272
Cites:
Cited by:
Related to: