15-418 / 15-618 Final Project

Parallel Multi-Agent Path Planning for Coordinated Robotic Manipulation

Team Members: Dylan Sun (bdsun), Jongsun Park (jongsunp)

Project Website

This website contains materials for our 15-418/15-618 final project.

Project Overview

We are going to study parallelization strategies for multi-agent path planning in a simplified robotic manipulation setting, where multiple fingers of a robotic hand must move from start nodes to designated goal nodes without colliding. Our original plan centered on CBS, but our current implementation has shifted toward a greedy repair planner because exact CBS was too expensive for the benchmark regime we want to study. We now have a working sequential baseline, an OpenMP top-K parallel version, benchmark infrastructure, and SVG visualization for comparing behavior and scalability.

Current Status and Revised Schedule

Completed so far: benchmark generation, collision detection, space-time A*, greedy repair baseline, OpenMP top-K repair evaluation, correctness tests, benchmark text output, and SVG visualizations. The main change from the proposal is that we are no longer using traditional CBS as the active baseline, because it was not practical on our target instances.