The paper proposes a database for multi-agent systems, the Multi-Agent Coordination Skill Database, that stores and utilizes past experiences to adapt to new collaborative tasks. The database stores coordinated behaviors with a unique skill representation captured by a Transformer-based skill encoder. The database can train the policy using a dataset augmented with retrieved demonstrations. The method was found to be effective in push manipulation tasks. The database is expected to be a valuable resource for policy learning.
Publication date: 5 Dec 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2312.02008
