Learning for TAMP using RAMP

Jack Collins, Mark Robson, Jun Yamada, Mohan Sridharan, Karol Janik, Ingmar Posner

Keywords: Benchmark, Task and Motion Planning, Manipulation, Simulation

TL;DR: RAMP is a benchmark for evaluating autonomous assembly and planning that is designed to be accessible, open-ended, and ideal for training and assessing learning for task and motion planning.

Abstract: RAMP, a benchmark for evaluating autonomous assembly and planning, provides an ideal environment for comparing approaches and assessing progress in Task and Motion Planning. RAMP is a recently released benchmark inspired by real-world industrial assembly tasks with the aim being to assemble a set of beams into predefined goal configurations. The benchmark is designed to be accessible with the physical benchmark created from 3D printed and other easily obtainable parts. The 3D printed parts can be extended upon and reconfigured into numerous designs, thereby making the benchmark open-ended. RAMP features a simulation environment to allow for virtual progress and therefore readily lends itself to learning-based approaches. To reduce the barrier to entry the benchmark includes a baseline and other conveniences, such as fiducial markers and layout templates, so users can focus on individual sub-tasks of the overall assembly challenge. RAMP has been designed intentionally to assess long-horizon planning and features many components that make it an ideal environment for learning and assessing task and motion planning.