Supplementary Video 1. Imitation rollouts of the expert imitator.
Prior Rollouts
Supplementary Video 2. Behavioral diversity yielded by rolling out the behavior prior with low sampling noise.Supplementary Video 3. Limited behavioral diversity yielded by rolling out the motion prior with low sampling noise.Supplementary Video 4. Increase in behavioral diversity — and resulting decrease in motion realism — yielded by rolling out the motion prior with high sampling noise.
Locomotion
Supplementary Video 5. SCAMPER exhibiting a walking gait during the locomotion task (target speed: 30 cm/s).Supplementary Video 6. SCAMPER exhibiting a trotting gait during the locomotion task (target speed: 50 cm/s).Supplementary Video 7. SCAMPER exhibiting a galloping gait during the locomotion task (target speed: 70 cm/s).Supplementary Video 8. SCAMPER exhibiting a galloping gait during the locomotion task (target speed: 120 cm/s).Supplementary Video 9. MIMIC-MJX exhibiting a walking gait during the locomotion task (target speed: 30 cm/s).Supplementary Video 10. Inability of MIMIC-MJX to generalize to a trotting gait during the locomotion task (target speed: 50 cm/s).Supplementary Video 11. Inability of MIMIC-MJX to generalize to a galloping gait during the locomotion task (target speed: 70 cm/s).Supplementary Video 12. Inability of the from-scratch model to perform a walking gait during the locomotion task (target speed: 30 cm/s).Supplementary Video 13. Inability of the from-scratch model to perform a trotting gait during the locomotion task (target speed: 50 cm/s).Supplementary Video 14. Inability of the from-scratch model to perform a galloping gait during the locomotion task (target speed: 70 cm/s).
Rearing
Supplementary Video 15. SCAMPER successfully completing the walk-then-rear task.Supplementary Video 16. Inability of MIMIC-MJX to complete the walk-then-rear task.Supplementary Video 17. Inability of the from-scratch model to complete the walk-then-rear task.
Arena Foraging
Supplementary Video 18. SCAMPER (early training): the model relies on stochastic sampling of its behavior prior to explore the foraging task structure.Supplementary Video 19. SCAMPER (mid training): as the model begins to receive reward signals from successful treat collection during the foraging task, it rapidly switches between stochastic sampling of the behavior prior and residual control of its motion prior.Supplementary Video 20. SCAMPER (fully trained): the model relies mostly on residual control of its motion prior to exploit the learned foraging task structure.Supplementary Video 21. MIMIC-MJX (fully trained): the model struggles to move efficiently and cannot collect all the treats within the allotted time in the foraging task.Supplementary Video 22. From-scratch model (fully trained): the model collects all the treats in the foraging task but does so with deeply unnatural motion.