Transfer Experiments

On this page, we present ten consecutive trials of our visually guided policy. Gaps are width 8-10cm and the robot starting location is varied from run to run. We deploy our policy trained in simulation on the real robot without any dynamics randomization or detailed environmental modeling.


We consider this transfer good, but imperfect. The gap-crossing task requires centimeter-level foot-placement accuracy for guaranteed success, which is sometimes denied due to a combination of sensor and actuator noise.

run0_reg.mp4
run9_reg.mp4
run7_reg.mp4
run6_reg.mp4
run4_reg.mp4
run2_reg.mp4
run5_reg.mp4
run1_reg.mp4
run8_reg.mp4
run3_reg.mp4

Across ten consecutive trials, we count:

  • Three (30%) complete successes.

  • Five (50%) single-foot slips; one foot slips into the gap but the robot remains stable to the other side.

  • Two (20%) multi-foot failures; multiple feet enter the gap, the robot loses stability.