LLM-Guided Adaptive Gait Control for Humanoid Robots via Multi-Modal State Reasoning

Abstract

Adaptive gait control for humanoid robots operating in unstructured real-world environments remains a critical open problem. Existing approaches, including classical model-predictive controllers and reinforcement-learning-based policies, exhibit limited cross-terrain generalisation and require costly retraining when encountering novel locomotion contexts. This paper presents LLM-GAC, an LLM-guided adaptive gait control framework that leverages large language models as high-level policy reasoners over fused multi-modal state representations drawn from inertial, vision, and force-torque sensing. The LLM encodes environmental context from natural-language task descriptions and robot state embeddings, generating structured gait tokens that modulate low-level locomotion controllers in real time. Experiments on a simulated Unitree G1 platform across five terrain categories demonstrate that LLM-GAC achieves a 92.3% gait stability score and a 2.1% fall rate, outperforming RL-PPO by 10.8 percentage points and MPC by 18.1 percentage points on stability, while reducing energy consumption to $124.6 \mathrm{W} \cdot \mathrm{s} / \mathrm{m}$. Mean adaptation latency is 38.6 ms, satisfying real-time constraints. Ablation studies confirm the contribution of each sensing modality, and cross-terrain success rates validate the generalisation capacity of the proposed framework.