Benchmarking and Enhancing Skill-Level Memory for Partially Observable Robotic Manipulation
The paper tests whether robot policies can remember hidden task state, then adds memory to improve execution under partial observability.
The authors introduce HIDE, a 15-task benchmark for manipulation settings where the same current observation can require different actions depending on earlier events. The tasks cover repetition counting, historical-state recall, and execution-progress tracking, with randomized starts and decision points. They also propose SEEK, a framework that combines three memory mechanisms for retaining evidence and tracking state. Existing policies perform poorly on HIDE, while memory augmentation improves success in simulation and real-world experiments. HF Daily Papers' note
The authors introduce HIDE, a 15-task benchmark for manipulation settings where the same current observation can require different actions depending on earlier events. The tasks cover repetition counting, historical-state recall, and execution-progress tracking, with randomized starts and decision points. They also propose SEEK, a framework that combines three memory mechanisms for retaining evidence and tracking state. Existing policies perform poorly on HIDE, while memory augmentation improves success in simulation and real-world experiments. HF Daily Papers' note
score 4