Finite-Horizon Performance Estimation for Online Parameter Refinement in RL–MPC Quadrupedal Locomotion
Year: 2026
Type: article
Abstract: Reinforcement learning integrated with model predictive control (RL-MPC) enables adaptive quadrupedal locomotion, but learned controller parameters do not explicitly account for their subsequent closed-loop performance. This study proposes a finite-horizon performance estimator (FHPE) to guide online parameter refinement. Trained on rollout data, FHPE predicts the finite-horizon performance of local parameter candidates to guide constrained refinement of the nominal RL proposal. The refined parameters are implemented through MPC and Adaptive Impedance Modulation (AIM), while Transient Residual Compensation (TRC) regulates joint-level responses through spectral regularization and risk-dependent gating. Ablation studies show that FHPE-guided refinement reduces the combined performance cost by 12.68% compared with equal-budget search, while AIM and TRC improve execution performance and transient regulation. Robot performance evaluations demonstrate improved high-speed locomotion capability, an approximately 26% reduction in peak joint acceleration on uneven terrain, and an approximately 43% reduction in estimated mechanical work over 10 m on flat ground in hardware experiments. These findings suggest that finite-horizon performance prediction can improve online parameter selection, while joint-level adaptation helps translate refined parameters into effective locomotion control.
Source: Applied Sciences
DOI: 10.3390/app162211478
Institution: Wuhan University of Technology
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
Award: G3929361599
OpenAlex ID: W7236739447
