Event
Paper proposes using coding agents to automatically design robot spatial memory
1 reports · 1 sources
Qianwei Wang et al. propose in the arXiv paper "Spatial Memory by Agents, for Agents" to have coding agents automatically design the construction and retrieval schemes for robot spatial memory: under a fixed development budget, a Designer Agent writes and iteratively revises the relevant code, with feedback provided by a fixed Query Agent. During evaluation, the generated implementation is first frozen, then memory is built from held-out observation histories (without access to downstream questions), and scored on five task tracks using the same Query Agent. In three independent design searches, the Codex / GPT-6 Astra combination achieved an equal-weight average score of 44.23 across the five tracks, while the strongest baseline supporting all five tracks, ReMEmbR, scored 29.32, a relative improvement of 50.9%. The paper has been submitted to ICRA 2027.
Report timeline
- Study on Coding Agents Automatically Designing Robot Spatial Memory: Codex / GPT-6 Astra Five-Track Average Improves 50.9% Over ReMEmbR
The arXiv paper "Spatial Memory by Agents, for Agents" investigates letting coding agents automatically design robot spatial memory: under a fixed development budget, a Designer Agent writes and iteratively revises memory construction and retrieval code, with feedback from a fixed Query Agent; the resulting implementations are frozen and then used to build memory from held-out observation histories for evaluation.
arXiv cs.RO ·