Sponsored by
NeurIPS 2026 · Competition Track
Build a single HVAC control policy that generalizes across commercial buildings it has never seen, jointly minimizing energy, carbon, and discomfort. A step toward scaling ML-driven building optimization across the globe.
About the challenge
Buildings cause nearly 40% of carbon emissions in the U.S. Of that total, HVAC systems are responsible for roughly 40 to 60%, and account for about 15% of global energy consumption. Yet despite advances in AI, most buildings still run about as inefficiently as they did decades ago.
The biggest obstacle is the lack of policy transfer. Most building control systems, whether simulation-based or learned, only work well for the building they were developed for, and fail to generalize to others.
To break through this, we propose a challenge to develop building foundation models that adapt to new buildings with minimal configuration, so building optimization can finally scale globally.
Participants build control policies on a set of simulated buildings, then have those policies evaluated on an unseen set of buildings that vary in size, layout, weather, material properties, and HVAC system. Success here would mark a real leap for global sustainability and show what ML can do against climate change.
The shape of the problem
The same policy is deployed, unchanged, on buildings it has never encountered, inferring its strategy from the building description and live sensor stream it is handed at runtime.
The goal is not to optimize one controller for one building. It is to learn one policy that works across arbitrary floor plans, HVAC topologies, and climates.
At each 5-minute step the policy issues discrete on/off commands and continuous setpoints, and their dimensions change from building to building.
A regret-based reward forces a genuine tradeoff: condition aggressively and burn energy, or switch off and lose comfort. The objectives pull against each other.