Sampling methods for motion planning and control using learned spaces

Standard grant project in years 2026-2028 supported by the Czech Science Foundation (GAČR) under research project No. 26-22606S

Applicant: Vojta Vonásek

Main Goal:

Development of new methods for motion planning in constricted spaces; new methods for detecting homotopy classes of paths, new stochastic robot controllers; publications in impacted journals and international conferences.

Abstract:

The project is motivated by a scenario where an assistant robot helps people manipulate objects. Sampling-based motion planning and control are challenging in this context, as the robot has many degrees of freedom and operates in constricted spaces (amongst or even in furniture, cupboards, etc.), which leads to the well-known "narrow passage" problem.

We aim to cope with this issue by constructing and exploring a lower-dimensional space (latent space) of the configuration space. We employ an autoencoder neural network to obtain the space and design new methods for its exploration. Searching in this space will enable faster and more efficient planning, particularly in constricted environments. We will also investigate the detection of homotopy classes to improve planning diversity and robustness. For control, we focus on developing new stochastic controllers based on Model Predictive Path Integral (MPPI).

Gallery:

(Content to be added)

In Media:

(Content to be added)

Publications:

(Content to be added)