Prorok Lab has three papers accepted at #ICLR2026! We will present work that spans robot policy watermarking, higher-order interactions in multi-agent pathfinding, and the role of diversity in MARL. Here’s the lineup:
📄 Remotely Detectable Robot Policy Watermarking by Michael Amir*, Manon Flageat*, and Amanda Prorok
How do you verify what policy a robot is deploying without access to its internals? CoNoCo embeds a spectral watermark into a robot’s motions via colored noise. The surprising part: it’s detectable from ordinary video footage alone, even through noisy, asynchronous observations—and without degrading policy performance. This opens the door to remote policy auditing for safety, accountability, and IP protection. https://lnkd.in/ei35uJZr
📄 Pairwise is Not Enough: Hypergraph Neural Networks for Multi-Agent Pathfinding by Rishabh Jain, Keisuke Okumura, Michael Amir, Pietro Lio’, and Amanda Prorok
Standard GNNs model pairwise interactions, but multi-agent pathfinding is fundamentally a group coordination problem. We propose HMAGAT, a hypergraph attention network that captures higher-order interactions, outperforming models 85× its size trained on 100× more data in dense environments. https://lnkd.in/eDbuAFzU
📄 When Is Diversity Rewarded in Cooperative Multi-Agent Learning? by Michael Amir*, Matteo Bettini*, Amanda Prorok
We know how to train heterogeneous agents, but when does heterogeneity actually improve performance? In cooperative multi-agent RL, we show this question is intricately related to the curvature of the reward function, and introduce HetGPS, an algorithm that automatically discovers environments where diversity pays off. https://lnkd.in/ejFUiwdB
Attending #ICLR2026? Come chat with us at our posters!
#Robotics #MultiAgentSystems #MachineLearning #ProrokLab