Learning Locomotion on Discrete Terrain via Minimal Proximity Sensing

[00:00:06]In this work, we modified the feet of  a commercially available animal D robot to house an STM Time of Flight proximity sensor.  The sensor is placed inside the rubber ball foot, with a machined aperture providing an  unobstructed field of view. Dedicated signal-processing electronics are  integrated within the foot assembly.

[00:00:31]This sensor divides its field  of view into a 4-by-4 grid, providing 16 sensing zones. By scanning the  terrain beneath the foot prior to contact, the sensor provides local geometric awareness  independent of global line-of-sight constraints.

[00:00:52]Leveraging foot-centric proximity sensor  feedback, we propose a minimal approach for learning robust locomotion  policies over discrete terrain.

[00:01:03]We simulate the large FoV  of the beams by ray casting at a higher resolution to approximate  the sensor reading in each grid cell.

[00:01:12]The resulting policy demonstrates  reliable traversal across multiple terrain types with a balance  of stability and agility.

[00:01:25]As benchmarks for comparison, we investigated whether imminent contact information could  be effectively recovered from conventional, base-centric sensors as an alternative  to the foot-centric proximity pipeline.

[00:01:39]We discovered that base-centric sensing  models achieve reasonable performance in absence of injected degradations. However, once  perturbations were introduced, the performance of base-centric sensing deteriorated markedly  faster than that of foot-centric sensing.

[00:02:02]We finally validate the proposed sensing framework  through extensive real-world experiments. Here we present five representative stepping-stone  scenarios, each beginning with an initial gap.

[00:03:01]These experiments showcased reliable  traversal of discontinuous terrain.

[00:03:08]Thank you for watching.