Temporal Multimodal Encoding for Reinforcement Learning-Driven Quadruped Robotic Control

In recent years, reinforcement learning (RL) has made significant progress in the field of legged motion control. Unlike traditional methods that rely on precise parameters, vision-free strategies based on RL extract features from a robot’s proprioceptive information through an encoder, enabling stable motion without visual input. However, multilayer perceptron (MLP) is predominantly utilized in existing encoders, and the static feature extraction mechanisms of MLP face difficulties in capturing dynamic characteristics in time-series data, consequently limiting the robot’s adaptability in complex environments. To tackle this problem, a temporal multimodal encoding (TME) method is proposed. In this method, the advantages of gated recurrent unit (GRU) and MLP are combined, facilitating the deep fusion of temporal and spatial features. Additionally, a contrastive learning mechanism is introduced, where external modalities are contrastively learned alongside features extracted from the robot’s historical proprioceptive information, thus improving the model’s representation capability and generalization performance. The effectiveness of the proposed method has been verified through simulation experiments on a quadruped robot.