Benchmarking and Enhancing Skill-Level Memory for Partially Observable Robotic Manipulation

The authors introduce HIDE, a benchmark comprising 15 tasks to evaluate manipulation memory under partial observability, covering historical-state recall and progress tracking. They also propose SEEK, a framework combining three complementary memory mechanisms. Evaluations show that existing policies struggle on HIDE, while memory augmentation improves task success in simulation and real-world experiments. Individual mechanisms benefit some tasks but can degrade others, highlighting the need to match memory design to task-specific requirements.

HIDE comprises 15 tasks covering repetition counting, historical-state recall, and execution-progress tracking.

SEEK's combination of memory mechanisms achieves the highest average success rate on HIDE among evaluated configurations.

Memory augmentation improves task success in both simulation and real-world experiments.