Student Policy Deployment

This page shows examples of the student DIC policy using real-time depth images and onboard state estimation to traverse gaps of variable spacing and width. We also present two failure cases that illustrate the limitations of our approach.

Success Cases

twogaps_se_vision.mp4

Two 10cm Gaps

vlc-record-2021-07-19-10h04m04s-GH010547.MP4-.mp4

Three 8cm Gaps

onegap_pronk_se_vision.mp4

One 15cm Gap

vlc-record-2021-07-19-10h05m15s-GH010541.MP4-.mp4

Three 15cm Gaps

Adaptive behavior in the student policy while pronking across 15cm gap. The commanded and actuated velocities during gap crossing (orange) deviate from those for flat ground (grey) after a gap has been perceived (top row).

Failure Cases

vlc-record-2021-07-19-10h03m08s-GH010549.MP4-.mp4

Rear Right foot slips off the near edge of the final gap.

vlc-record-2021-07-19-10h04m50s-GH010544.MP4-.mp4

Front feet make contact on the platform but slip due to insufficient contact force. This is a common symptom of an inaccurate body height estimate.

lllustration of State Estimate Drift

Body and foot position tracking using onboard state estimator (IMU + Kinematics + Kalman Filter). To cross gaps with the onboard state estimator, the high-level policy must compensate for substantial drift via closed-loop incorporation of vision.

Body and foot position tracking using the motion capture system. The low-level controller closely tracks the ground-truth desired pose.