Abstract: Reinforcement-learning-based controllers for lightweight robot control in structured, rather than complex, scenarios have received comparatively little attention. Lightweight control aims to achieve performance comparable to full observationbased control by relying on a minimal set of key observation factors. This study finds that different key observation factors become important under different commands. Therefore, the effectiveness of such methods still requires further enhancement. To better understand the influence of observation states on action generation and enhance lightweight control performance, this paper proposes a method based on an improved Integrated Gradients algorithm to quantify the importance of observation variables in quadruped robot control. By combining sampling and fitting techniques, the variation pattern of observation importance weights is derived with respect to different velocity commands, and the optimal observation subsets are identified under specific command conditions. To enable real-time switching of observation subsets during inference based on command input, a multi-encoder training framework is introduced. This framework encodes and learns multiple observation subsets using a shared replay buffer and a Multi-stage Encoder Switching Architecture. Simulation results demonstrate that different observation subsets are optimal for different command ranges. The proposed command-driven observation selection strategy improves the overall performance score by 5.33% compared with fixed Determined State Observation configurations. Real-world experiments validate similar trends, highlighting the potential of this approach for lightweight robot control and the exploration of biologically inspired locomotion mechanisms.

Key words: Reinforcement learning')">Reinforcement learning, Legged robots, Observation selection, Privileged learning