New AI Trick Teaches Robot Teams to Cooperate on the Cheap

What if a single AI could learn a task on its own, then be taught to work in a team without ever retraining the brain it already has? That's the idea behind MATES — Multi-Agent Observation Transformation for Existing Single-Agent Policies — an input-side adaptation framework from researchers Elie Abboud and Oren Gal. From multi-agent experience, MATES learns a small adapter that maps a multi-agent observation into the format a frozen single-agent policy expects, producing actions suited to the shared environment while the pretrained policy itself stays untouched, architecture and all. The underlying MARL algorithm's objectives and update procedures are retained. Testing MATES with both on- and off-policy algorithms across lifelong pathfinding, navigation, and cooperative discovery, spanning discrete and continuous observation and action spaces, the results were striking: MATES optimized only 3.5–7.3% as many parameters as full-policy training, yet consistently outperformed MARL training from scratch across all evaluated settings. It approached the performance of full fine-tuning, stayed competitive overall with demonstration-based baselines, and kept strong task performance even at team sizes it never encountered during training. The takeaway: under this observation structure, effective multi-agent behavior can be learned without modifying the policy that encodes individual competence.