Reproduce “SpaceAct: Benchmarking 3D Spatial and Temporal Reasoning via Closed-Loop Action and Active Perception” 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
Chenglong Shi, Zekai Yuan, Yue Song, Jon Pu, Will Wu
SpaceAct tests models’ ability to infer 3D geometry, gather views, and execute actions under contact and timing constraints. It checks swept paths and returns collision feedback. Across 22 episodes on eight L5 tasks, three models averaged scores of 65.1 unaided, 80.3 with a plan, and 96.6 with a plan plus exact state; collisions fell from 5.2 to 0.5 with state.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.