Reproduce “GeoMind: A Benchmark for Spatial Understanding in Robotic Manipulation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jinghe Wang, Xinrui Cao, Duo Wu, Chenghao Gu, Yong Zhong, Linjia Kang, Tianyi Xiong, Zhi Wang

GeoMind evaluates spatial reasoning and memory in tabletop manipulation by separating choosing targets from current scene relations from preserving object identity after rearrangement. It uses verified simulated tasks with varied layouts and deterministic target grounding. The best learned policy achieved 26.4% task success, versus 99.3% for a simulator-state oracle; results are limited to frozen policies in simulation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.