Event
LogLook proposes execution-log-style memory for robot planners
1 reports · 1 sources
Tianyi Jiang, Tianyao Tong, and Zhenshan Bing submitted a paper proposing an execution-log-style memory method called LogLook for frontier vision-language model robot planners. The method has the system retain execution logs and the model's prior notes, with the model deciding which image frames to review: returning frames only where expectations change, first skimming the demonstration and then inspecting selected parts. Under the benchmark's built-in motion planner, LogLook's success rate is comparable to note-only memory, but with 43% fewer tokens and less than half the image usage, and notably fewer failures caused by memory or observation errors. Under the direct-control setting, LogLook solves 93.5% on the RoboMME test set. All the above numbers are results reported in the paper.
Report timeline
- LogLook: Execution-Log Memory for Frontier-Model Robot Planners
LogLook proposes an execution-log-style memory that lets frontier vision-language model planners record executed actions and earlier notes, and selectively review image frames based on expected changes. With the benchmark's built-in motion planner, it matches the success rate of note-only memory while using 43% fewer tokens and less than half the images, and it causes notably fewer failures due to memory or observation errors; under direct control, it solves 93.5% of the RoboMME test set.
arXiv cs.RO ·