Motion Planning on Visual Manifolds

TL;DR

This thesis introduces Visual Configuration Space (VCS), enabling robots and agents to learn body structure and plan obstacle-free motions solely from visual data without prior geometric knowledge.

Contribution

It presents a novel geometry-free approach to motion planning using visual manifolds, applicable to robots, human learning, and digital avatar animation.

Findings

VCS allows obstacle-free motion planning from images alone.

Demonstrates learning body structure without geometric models.

Generates natural head motions for virtual avatars.

Abstract

In this thesis, we propose an alternative characterization of the notion of Configuration Space, which we call Visual Configuration Space (VCS). This new characterization allows an embodied agent (e.g., a robot) to discover its own body structure and plan obstacle-free motions in its peripersonal space using a set of its own images in random poses. Here, we do not assume any knowledge of geometry of the agent, obstacles or the environment. We demonstrate the usefulness of VCS in (a) building and working with geometry-free models for robot motion planning, (b) explaining how a human baby might learn to reach objects in its peripersonal space through motor babbling, and (c) automatically generating natural looking head motion animations for digital avatars in virtual environments. This work is based on the formalism of manifolds and manifold learning using the agent's images and hence we call it Motion Planning on Visual Manifolds.