Spot Whole-Body Reinforcement Learning
Learned box tipping, crouching, and body reorientation.
Manipulation
Tip the box.
A learned policy tips the box. Three successful attempts, with episode resets retained.
Visual description
Spot reaches toward a yellow box with its arm and tips it onto its side. The scene resets between three successful tipping attempts. The studio render replays recorded simulator states from a saved policy.
Whole-body control
Get low. Stay balanced.
Lowering toward a 24 cm body-height target, with the arm stowed.
Visual description
Spot bends its legs and lowers its body while its arm remains folded. A readout shows actual and target body height. The excerpt follows a descent from standing toward a 24-centimeter height target and the beginning of the low hold.
The outtake
Almost had it.
The box wins this round. No stable tip before the timeout.
Visual description
Spot repeatedly reaches and adjusts its arm around the box. The box remains upright through the end of the saved episode. This attempt is unsuccessful and is included as an outtake.
About these rollouts
Native HD studio replays of recorded simulator states from separate reinforcement-learning studies. The saved policies are frozen during state capture; the studio replays their motion kinematically. Box-tipping footage comes from a partial policy checkpoint. The crouch policy learns leg-joint control with the arm stowed; its full assessment did not meet the study's criterion in every tested condition. Selected behaviors do not estimate overall success or demonstrate physical-robot performance.
Simulation and rendering: NVIDIA Isaac Lab. Spot model and configuration: ReLIC.
Media sources & edit record · Related: Contact IK · SpotNLP