Excited to share that our paper, “Scaling Multi-Agent Environment Co-Design with Diffusion Models,” has been accepted at ICML 2026! 🎉
With Hao Xiang Li, Michael Amir, Amanda Prorok
Engineers train robots to pick and place boxes. Architects design warehouse layouts.
But these two problems are usually treated separately. In this work, we ask what happens when we design the robots’ behavior and the environment together.
We use diffusion models to generate better environments, guided by what the agents learn during training. As the agents improve, they provide a stronger signal for which environments are useful; as the environments improve, they create better training settings for the agents.
This makes it possible to scale multi-agent co-design to much larger problems, including warehouse layouts and wind farm control.
Paper: https://lnkd.in/eEjtQrMr
#ICML2026 #MachineLearning #Robotics #DiffusionModels #MultiAgentSystems