Hierarchical meta-reinforcement learning via automated macro-action discovery

Minjae Cho, Chuangchuang Sun · 2026

Read Hierarchical meta-reinforcement learning via automated macro-action discovery by Minjae Cho, Chuangchuang Sun (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/PMC13572510.txt.

Description

Minjae Cho and Chuangchuang Sun present a 2026 method that helps robot learning agents pick up new tasks fast. Agents that learn by trial and error often forget old skills when a new task clashes with them. The method stacks three levels: one reads the task, one finds broad moves that fit many tasks, and one picks fine actions.

The broad moves, called macro-actions, are found on their own and stripped of task-specific detail, so they can be reused across jobs. They point the fine policy toward goal states and ease the load on each level. Each level trains on its own tailored scheme, which keeps the stack stable. On the MetaWorld test suite, the method learned with fewer trials and won more often than past leaders. Code is posted online for others to check.

Catalog record

CREATORMinjae Cho, Chuangchuang Sun
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE92285
LICENSEhttps://creativecommons.org/licenses/by/4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/article/MED/PMC13572510

Browse the full library · About the Tower of Babel