.png){kind=icon}
Return
Trajectory Segmentation and Self-Supervised Classification for Subtask Discovery
.png){kind=icon}
PRE
AI
DOI：10.1007/978-981-95-4972-6_8.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Discovering and solving subtasks individually is a common approach to tackle long-horizon tasks. However, the existing approaches for subtask discovery primarily aim to accelerate downstream task training and often fail to identify semantically meaningful subtasks. We aim to discover semantically meaningful subtasks, in a robotic dataset where the demonstrations complete subtasks in varying orders, to enable sequential execution of subtask policies. We first segment the demonstrations into subtask working and transition phases in a semantically meaningful way. We then employ a self-supervised approach that is agnostic to the number of labels to classify the trajectory segments to obtain distinct subtasks. We trained a set of independent policies solely on the corresponding subtask demonstration segments. The policy learns to decide when to switch to the next subtask policy. Our experiment showed that executing these subtask policies sequentially achieved 98% of the total rewards of a baseline policy trained on complete demonstrations. This result demonstrates that our method identifies semantically meaningful subtasks with accurate boundaries, ensuring the seamless execution of multiple independent subtask policies.
Keywords:
Robot Manipulation
Behaviour Cloning
Subtask Discovery
Self-supervised classification
Journal
A
IF:
0
Papers:
31
Citations:
0
Organization
Cited Papers
.png){kind=icon}
ShareSaveShareSaveShareSave
no more
no more