March 27, 2025
When attempting to find a behavior to satisfy an objective on a robotic system, the first step is defining that objective. While tasks can be encoded in natural language or via a manually defined reward function in Reinforcement Learning (RL), these methods typically carry biases and can lead to unintended interpretations or ambiguities. Specification-guided learning, which uses formal logic (often temporal) to guide learning algorithms, offers a solution to these challenges. This approach is particularly relevant for multi-agent (MA) systems, where multiple agents interact in complex environments like swarms of drones or autonomous vehicle fleets. Some key challenges in applying Specification-guided learning to MA systems are a reliance on the global knowledge of the system and struggles with scalability in the number of agents when adding safety constraints. To address these issues, we separate the objective-achieving process from collision-avoidance in our temporal logic tasks, leveraging existing decentralized MA collision avoidance methods (viz. GCBF+) to achieve Signal Temporal Logic (STL) tasks under specific assumptions, such as known dynamics and homogeneous tasks. Finally, we propose methods to relax some of these assumptions (particularly homogeneous tasks) by enabling agents to achieve different objectives without retraining through the multi-modal modeling capabilities of generative algorithms such as diffusion models.
About Joe Eappen
Joe is a PhD student at Purdue University under Prof. Suresh Jagannathan, focusing on specification-guided learning for multi-agent systems. His studies have produced publications at CoRL and ECML, with collaborations featured at ICML, IROS, ICRA, and in RA-L. Joe has also interned at JPMorgan Chase & Co. where he worked on Offline Reinforcement Learning (RL) and Synopsys where he worked on Graphical Neural Networks (GNNs). His research emphasizes temporal logic constraints in robotic and multi-agent systems, enabling safer and more reliable control.