Test-Time Multi-Agent Tool Use: Collaborative Information Acquisition in Dynamic Environments

Most test-time adaptation research focuses on single-agent scenarios. However, studies of primate tool use (Bandini, 2020) and collaborative human-robot settings suggest that social and collective learning can dramatically accelerate adaptation and innovation. This project would develop a system where, at test time, multiple agents can observe, mimic, or adaptively diverge from each other's tool use strategies, potentially sharing cues or outcomes. This could be implemented in robotics, where a fleet of manipulators share information about failed/successful grasps (see Mosbach & Behnke, 2023), or in multi-user digital systems (inspired by user-level anomaly detection in Parres-Peredo et al., 2019). The novelty is in treating test-time adaptation as a distributed, social process—exploring how coordinated tool use and information sharing can uncover solutions that would be invisible to isolated learners, especially in non-stationary or adversarial environments.

References:

1. Learning Generalizable Tool Use with Non-Rigid Grasp-Pose Registration. Malte Mosbach, Sven Behnke (2023). 2023 IEEE 19th International Conference on Automation Science and Engineering (CASE).
2. Implementing long‐term baselines into primate tool‐use studies. E. Bandini (2020). American Journal of Primatology.
3. Unexpected-Behavior Detection Using TopK Rankings for Cybersecurity. A. Parres-Peredo, I. Piza-Dávila, Francisco Cervantes (2019). Applied Sciences.