Open Collections

Khoshsiyar, Niloofar

University of British Columbia

We propose a contact-centred approach for synthesizing physics-based character animation. Rather than using target keyframe poses to help guide a motion, our framework uses sparse keyframes based on a small number of pairwise contacts between the body and the environment or between two parts of the body. These serve as loose specifications of motion strategies, which can then be solved online at interactive rates, using a model predictive control (MPC) framework. This allows for motion generalization across environment variations as well as producing diverse output motions due to the stochastic nature of the MPC solver. We demonstrate the advantage of this method in contrast to deep reinforcement learning methods which can require hours or days to compute a new control policy after modifications to the reward functions or the environment properties. Moreover, our framework exploits partwise motion planning results when this proves beneficial, i.e. in less coordinated tasks, while defaulting to whole-body motion planning when that proves to be more effective, i.e. in motions that require full-body coordination. Our results present a range of challenging contact-rich motions and everyday tasks that can be synthesized using this framework, showcasing both motion diversity and generalization capabilities.

Text

2024-03-15

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/87587

Computer Science

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/