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

TL;DR

Look Forward to Walk Backward (LF2WB) is presented, 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.

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.

Figures (4)

• Figure 1: We deploy our policy in real-world environments, demonstrating agile forward and backward motion over discrete terrains that combine steps and gaps. The robot memorizes the surrounding terrain during forward motion and retrieves useful information from the memory state when moving backward, enabling highly dynamic backward locomotion using only a forward-facing depth camera.
• Figure 2: Overview of the proposed LF2WB framework. We concurrently train a delta-rule estimator and an asymmetric actor-critic. The estimator learns a compact terrain memory with strong retrieval ability for backward motion, enabling the policy to execute dynamic backward maneuvers using only forward egocentric vision.
• Figure 3: RMSE of the body-centric elevation estimate along the executed path. The robot starts on the left of the image and walks backward to the right; $0\,\mathrm{s}$ corresponds to the leftmost position where the backward phase begins. Curve locations at terrain discontinuities are marked by arrows. Once the robot falls, we stop recording error.
• Figure 4: Real-world deployment on a robot equipped with a forward-facing depth camera. For clarity, all sequences are visualized such that the robot moves from left to right.