A hierarchical multi-encoder fusion architecture for robust quadrupedal locomotion

International Journal of Advanced Robotic Systems

卷:23

Original Research Article

Ling Tang¹  Yibo Zhao¹  Heyao Wu¹  Shimin Wei¹ 

[1] , Beijing, China

关键词: legged locomotion;    deep learning;    reinforcement learning;    robot sensing systems;    robustness;   

DOI  :  10.1177/17298806261477177

 received in 2025-12-19, accepted in 2026-07-15,  发布年份 2026

来源: Sage Journals

【 摘 要 】

Quadrupedal robots demonstrate impressive mobility in unstructured environments, yet achieving robust locomotion under limited sensing remains a significant challenge. Recent progress in deep reinforcement learning suggests that proprioceptive feedback alone can support adaptive gait generation without external perception. Building on this insight, we propose a hierarchical multi-encoder actor framework for proprioception-only quadrupedal locomotion. The architecture consists of three layers: a sensor-encoder layer that independently learns velocity, contact, and predictive proprioceptive representations; a fusion layer that integrates these latent features into a compact unified space optimized via PPO; and an action layer that produces motor commands from both the fused representations and current observations. This hierarchical design facilitates modular representation learning and implicit terrain inference, enabling stable and adaptive locomotion on unseen and challenging terrains—without relying on cameras or elevation maps.

【 授权许可】

CC BY-NC   
© The Author(s) 2026

【 图 表 】

Figure 18

Figure 16

Figure 9

Figure 10

Figure 9

Figure 4

Figure 3