Versatile motion priors for artist-directable physical characters

Motion priors that let artists direct a physical (robotic) character's movement while keeping it physically plausible — Bächer's bipedal-droid line.

An encoder is trained to generate latent representations of kinematic motion from motion data and a control policy is trained to generate dynamics informed output for a robotic system based on the latent representations of kinematic motion. The encoder and the control policy are deployed in the robotic system to generate the dynamics informed output based on the motion data.

FIELD (1) The present application relates to systems and methods of generating physics-based or robotic characters based on artist input. BACKGROUND (2) Conventional animation techniques may be difficult to apply to robotics, as animated characters do not always follow the laws of physics. Physical robots, in contrast, must observe the laws of physics to remain dynamically stable. However, it is often desirable (e.g., where autonomous robots are provided for specific characters) to utilize animation techniques to imitate the style of a particular character, such as the gait and movement characteristics of the character while performing tasks such as everyday tasks, athletic motions, and/or other types of movements. Existing physics-based character tools generally provide for only indirect control of a character, making integration with traditional animation tools difficult. Further, such tools often use motion capture data as input. Generally, motion capture data is not sufficient for autonomous character control, as motion sequences provide only predefined responses and do not allow characters to respond to uncertain conditions. Additionally, there may be a need to update animation or provide different robotic devices different animation or movement skills and conventional techniques are time intensive requiring much iteration and training before deployment. BRIEF SUMMARY (3) A computer implemented method includes training an encoder to generate latent representations of kin