Overview and Results

We learn a policy from raw visual inputs to trajectory parameters in simulation and achieve near-optimal gap-crossing performance under fixed contact schedules.

rss_adaptive_trot_graphic4_recording.mp4

Visually Guided Trot - Simulation

rss_blind_trot_graphic3.mp4

Blind Trot - Simulation

rss_adaptive_pronk_graphic3.mp4

Visually Guided Pronk - Simulation

rss_blind_pronk_graphics4_recording.mp4

Blind Pronk - Simulation

We transfer our policy to the real system without domain randomization or detailed environment modeling.

run0_reg.mp4

Visually Guided Trot - Deployment Success `A`

run3_reg.mp4

Visually Guided Trot - Deployment Success `B`

View ten consecutive runs here: Trotting Transfer Experiments

We demonstrate the extension of our learning framework to realistic, highly dynamic leaps across gaps up to 1.5x the quadruped body length.

rss_dynamic_jump_realtime2.mp4

Crossing Large Gaps in Simulation

rss_dynamic_jump_realtime_Slomo.mp4

Slow-Motion

Our learned policy can succeed in novel environments where low-level controller behavior remains robust.

run0_sheet.mp4

Deployment Robustness - Surface Friction

run0_mattress.mp4

Deployment Robustness - Surface Compliance

View analysis of a relevant baseline here: Baselines