Mobile Construction Benchmark

Wenyu Han, Chen Feng, Haoran Wu, Alexander Gao, Armand Jordana, Dongdong Liu, Lerrel Pinto, Ludovic Righetti

Keywords: localization, dynamic environment, deep reinforcement learning, benchmark

Abstract: We need intelligent robots to perform mobile construction, the process of moving in an environment and modifying its geometry according to a design plan. Without GPS or similar techniques, carefully engineered and learning-based methods face challenges to exactly achieve the plan due to the difficulty of accurately localizing the robot while strategically evolving the environment, because common tasks (manipulation/navigation) address at most one of the two coupled aspects. To seek a generic solution, we simplify mobile construction in 1/2/3D grid worlds to benchmark the performance of existing deep RL methods on this partially observable MDP. Our results show that the coupling makes this problem very challenging for model-free RL, and emphasize the need for novel task-specific solutions.

One-sentence Summary: Simultaneously localizing a robot while changing the environment is very difficult for model-free RL without task specific algorithms.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Reviewed Version (pdf): https://openreview.net/references/pdf?id=8MzocUFrMx