Open Collections

Ioannidis, Nicholas

University of British Columbia

Diffusion models can be used as a motion planner by sampling from a distribution of possible futures. However, the samples may not satisfy hard constraints that exist only implicitly in the training data, e.g., avoiding falls or not colliding with a wall. We propose learned viability filters that efficiently predict the future success of any given plan, i.e., diffusion sample, and thereby enforcing an implicit future-success constraint that favours plans likely to succeed in the future. These filters identify samples from the output of a diffusion-based motion planner with respect to implicit constraints and take the general form of a learned Q-function. This allows for efficient online planning with diffusion models, including for situations where the diffusion model and viability filter have different access to environment observations. Multiple viability filters can also be composed together so that they are each taken into consideration. We demonstrate the approach on detailed footstep planning for challenging 3D human locomotion tasks, showing the effectiveness of viability filters in performing online planning and control for box-climbing, step-over walls, and obstacle avoidance. We further show that using viability filters is significantly faster than guidance-based diffusion prediction.

Text

2025-04-28

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/90859

Computer Science

University of British Columbia

UBCV

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