Real-time Whole-body Model Predictive Control for Bipedal Locomotion with a Novel Kino-dynamic Model and Warm-start Method

TL;DR

This work addresses real-time WB-MPC for bipedal locomotion by introducing a kino-dynamic model that merges LIPFM with full-body kinematics and uses ZMP instead of contact wrenches, reducing computation and avoiding discontinuities during contact transitions. A modularized, lightweight MLP warm-start and a ZMP-based WBC are integrated to provide fast, reliable initial guesses and stable impulses, enabling real-time control across walking phases. Comparative analysis shows substantial latency reductions (~44% vs WB models) and superior robustness under perturbations, with sub-1 ms warm-start latency and fast convergence in simulations and real-robot tests (latency ≈ 16–17 ms, ~2–3 iterations). The results demonstrate practical real-time applicability of WB-MPC for bipedal walking, while acknowledging limitations in vertical COM motion and flight-phase dynamics for future improvements.

Abstract

Advancements in optimization solvers and computing power have led to growing interest in applying whole-body model predictive control (WB-MPC) to bipedal robots. However, the high degrees of freedom and inherent model complexity of bipedal robots pose significant challenges in achieving fast and stable control cycles for real-time performance. This paper introduces a novel kino-dynamic model and warm-start strategy for real-time WB-MPC in bipedal robots. Our proposed kino-dynamic model combines the linear inverted pendulum plus flywheel and full-body kinematics model. Unlike the conventional whole-body model that rely on the concept of contact wrenches, our model utilizes the zero-moment point (ZMP), reducing baseline computational costs and ensuring consistently low latency during contact state transitions. Additionally, a modularized multi-layer perceptron (MLP) based warm-start strategy is proposed, leveraging a lightweight neural network to provide a good initial guess for each control cycle. Furthermore, we present a ZMP-based whole-body controller (WBC) that extends the existing WBC for explicitly controlling impulses and ZMP, integrating it into the real-time WB-MPC framework. Through various comparative experiments, the proposed kino-dynamic model and warm-start strategy have been shown to outperform previous studies. Simulations and real robot experiments further validate that the proposed framework demonstrates robustness to perturbation and satisfies real-time control requirements during walking.

Figures (7)

• Figure 1: Proposed novel kino-dynamic model: Integration of LIPFM and FKM.
• Figure 2: The proposed control framework using novel kino-dynamic model.
• Figure 3: Modularized MLP for finding good initial guess
• Figure 4: Latency graph for 1 DDP iteration during walking
• Figure 5: The residual cost of WB-MPC under perturbation