Event
LSMM Spatial Memory Architecture for Legged Robot Maze Exploration
1 reports · 1 sources
Leon Magnus and co-authors propose LSMM, a local spatial matrix memory architecture based on Hadamard memory, which lets reinforcement learning agents store task-relevant information in position-based memory units. The goal is to address the problem that humanoid and legged robots in complex environments need to remember information that is only used long after it has left their field of view. On an autonomous maze exploration task, the authors report that LSMM's maze coverage is comparable to strong recurrent networks and limited-history baselines, and that it is the only architecture able to remember the maze layout, with reconstruction error reduced to one third. These results come from the team's own evaluation and have not yet been independently verified.
Report timeline
- LSMM: Learning Spatial Memory for End-to-End Legged Robot Exploration
The authors propose Local Spatial Matrix Memory (LSMM) based on Hadamard memory, which lets reinforcement learning agents store task-relevant information in position-based memory cells. On an autonomous maze exploration task, LSMM's maze coverage is comparable to strong recurrent networks and finite-history baselines, and it is the only architecture that remembers the maze layout, reducing reconstruction error to one third. The work targets the problem that humanoid and legged robots in complex environments need to remember information long after it leaves their field of view before it is used.
arXiv cs.RO ·