Spatial Memory by Agents, for Agents
Published in arXiv preprint arXiv:2610.22736, 2026
Recommended citation: Q. Wang, C. Zhang, Z. Luo, P. Su, S. Yang and H. Xu. Spatial Memory by Agents, for Agents. arXiv preprint arXiv:2610.22736, 2026. https://arxiv.org/abs/2610.22736
We let coding agents design a robot’s spatial memory: they write code, run benchmarks, and iterate on feedback to decide how a scene should be stored and how it should be retrieved when needed. We call this Guided System Improvement (GSI). Across five task tracks, the best configuration improves the average unified score by about 51% over the strongest all-task baseline. Joint work with Mondo Robotics.