Search papers, labs, and topics across Lattice.
The paper introduces Look Forward to Walk Backward (LF2WB), a novel terrain-memory locomotion framework for legged robots that enables collision-free backward locomotion using only forward-facing depth cameras. LF2WB creates a compact associative memory of the terrain during forward motion using a delta-rule selective update mechanism, and then retrieves this memory to guide backward movement. Experiments in simulation and real-world environments demonstrate that LF2WB significantly improves backward agility on complex terrains compared to purely proprioceptive controllers.
Legged robots can now nimbly walk backwards on challenging terrain using only a forward-facing camera, thanks to a clever memory system that recalls the path they already traversed.
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.
