Look Forward to Walk Backward: Efficient Terrain Memory for Backward Locomotion with Forward Vision

2026-03-03

AI Summary

This work addresses the challenge of terrain preview and obstacle avoidance during backward locomotion in quadrupedal robots, which is hindered by the absence of rear-facing vision and consequently compromises motion stability. To overcome this limitation, the authors propose a compact associative memory framework that leverages forward-facing depth sensing and proprioception: terrain information is encoded during forward traversal and later retrieved during backward motion to enable vision-free obstacle avoidance. A novel selective memory update mechanism based on the delta rule is introduced, which softly overwrites outdated entries within an active subspace, achieving efficient and low-overhead memory management. The system employs a hardware-efficient parallel training strategy and, at deployment, requires only a constant-size internal state to perform recursive inference in constant time. Simulations and real-world experiments demonstrate significant improvements in backward agility and obstacle avoidance on complex terrains, with compatibility for low-cost onboard processors.

Technology Category

• Intelligent Robots: Learning & Optimization for ROB
• Search and Optimization: Learning to Search
• Humans and AI: Human-Aware Planning and Behavior Prediction

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Legged robots with egocentric forward-facing depth cameras can couple exteroception and proprioception to achieve robust forward agility on complex terrain. When these robots walk backward, the forward-only field of view provides no preview. Purely proprioceptive controllers can remain stable on moderate ground when moving backward but cannot fully exploit the robot's capabilities on complex terrain and must collide with obstacles. We present Look Forward to Walk Backward (LF2WB), an efficient terrain-memory locomotion framework that uses forward egocentric depth and proprioception to write a compact associative memory during forward motion and to retrieve it for collision-free backward locomotion without rearward vision. The memory backbone employs a delta-rule selective update that softly removes then writes the memory state along the active subspace. Training uses hardware-efficient parallel computation, and deployment runs recurrent, constant-time per-step inference with a constant-size state, making the approach suitable for onboard processors on low-cost robots. Experiments in both simulations and real-world scenarios demonstrate the effectiveness of our method, improving backward agility across complex terrains under limited sensing.

Problem

• backward locomotion
• forward vision
• terrain memory
• legged robots
• limited sensing

Innovation

• terrain memory
• backward locomotion
• forward vision
• associative memory
• legged robots