Reproduce “Spatial-Interactor: Learning Spatial Reasoning through Interaction with the Observable Physical World” 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

Kaixiang Yao, Xu Wang, Miao Pan, Hu Xiyue, Jiaxin Li, Weishi Wang, Daniel Dahlmeier, Jintao Chen and 5 more

Spatial-Interactor trains models to reason about changing three-dimensional scenes by learning object and camera transitions from before-and-after observations and composing them across trajectories. It uses training-only traces for process distillation. In a matched Qwen2.5-VL-7B comparison, OPD scores 60.1 Overall versus 59.2 for GRPO, with larger gains in Route Planning and Camera Displacement. This supports the combined addition, not either component alone.

Summary by alphaXiv.

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