The Smart Buildings Challenge: Learning Foundation Models for Real-World Optimization at Scale

AI summary

Overview

Research area: Artificial intelligence for climate and sustainability — specifically reinforcement learning, transfer learning, and foundation models applied to commercial building HVAC control.

Technical level: Intermediate. The paper is a competition proposal rather than an empirical results paper; it defines an environment, API, and scoring rule, and states that no domain knowledge is required to participate.

Scope (one sentence): The paper proposes a Kaggle-hosted NeurIPS 2026 competition in which participants train a single control policy on 6 simulated buildings and are scored on 6 unseen buildings at 5 unseen weather stations, testing whether a "building foundation model" can transfer across buildings and climates.

What This Paper Is About

Buildings — and especially their heating, ventilation, and air conditioning (HVAC) systems — consume enormous amounts of energy, yet most buildings still run on inefficient control logic because optimized controllers do not transfer from one building to another. The authors argue that the bottleneck is not optimization algorithms but transfer: policies are tied to a specific building management system, weather pattern, HVAC configuration, or data ontology, and cross-building training data is scarce. Their response is to launch a competition that explicitly rewards generalization — training on one set of simulated buildings and being scored on a held-out set of buildings and climates never seen before.

Key Contributions

1. A competition framing built around transfer, not single-building tuning. To the authors' knowledge, no previous competition has focused on foundation models for buildings — creating digital twin models that transfer from building to building. They identify this as "arguably the largest unsolved problem in this entire domain of research."

2. A procedural, unbounded building generator. Built on the HVAC simulation platform from the Smart Buildings Control Suite with community improvements, the generator produces an infinite set of building simulators. It varies floors (2–4), width and length, perimeter shape, room layout, occupancy, envelope construction (six material properties), and the number of Air Handler Units (AHUs) and the zones they serve (at most 10 zones per AHU), with a single hot-water system per building and one VAV box per zone. Each building is exposed as a Gym-compatible RL environment.

3. A strict held-out evaluation protocol with documented anti-leakage measures. Participants receive 6 training buildings each paired with one weather strip from each of five climate stations (30 training scenarios), and are evaluated on 6 unseen buildings at 5 climate stations that appear nowhere in training. The 6 held-out building configurations are generated from random seeds known only to the Organizers and stored in a private repository; no artifact derived from them is released until the competition closes.

4. A regret-based, bounded reward and an accompanying baseline suite. The leaderboard uses a weighted reward over comfort, energy cost, and carbon emissions, scored over a 30-day, 5-minute-timestep episode, with four hand-written reference controllers shipped in the starter kit for comparison.

Main Findings

• Transfer is the core obstacle, not optimization. Most control policies — simulation- or ML-based — fail to generalize across buildings for two reasons the authors name: system-specific dependencies (solutions tied to a particular building management system, weather pattern, HVAC system, or data ontology) and lack of transfer training (controllers trained on a single building because cross-building data is scarce).

• The evaluation is a pure held-out transfer test. Training covers 6 buildings × 5 climate stations = 30 development scenarios; evaluation covers 6 unseen buildings × 5 unseen stations = 30 scored rollouts per submission. Transfer is therefore required across both buildings and climates.

• The reward is bounded above by zero. Each 5-minute step yields r_t = w_p·p_t − w_c·c_t − w_e·e_t with (w_p, w_c, w_e) = (0.6, 0.2, 0.2), where p_t is occupant comfort (productivity), c_t is energy cost (electricity and gas), and e_t is carbon emission. All three terms are penalties. Comfort is a regret term: 0 when every occupied zone is inside its comfort band, and negative otherwise without bound — falling linearly up to 5 K outside the band and quadratically and unboundedly beyond. The largest per-step reward observed is exactly 0, attained at 95 steps, all of them with the building unoccupied and the plant idle.

• Comfort is weighted by occupancy, not floor area. The comfort term is summed over zones weighted by each zone's own occupancy, so a single badly conditioned zone can dominate an episode.

• The leaderboard metric is mean absolute error against zero. With 𝒮 the set of evaluation scenarios, MAE(π) = −(1/|𝒮|) Σ R̄⁽ˢ⁾(π), where R̄⁽ˢ⁾(π) = R⁽ˢ⁾(π)/T and T = 8640. Lower is better. Because r_t ≤ 0, minimizing MAE is identical to maximizing return on every individual scenario, and the headline score is the plain mean over all 30 scenarios with no grouping by building or weather.

• The simulator is deterministic. The same policy on the same scenario reproduces the same return bit for bit across processes and machines, so each (building, weather) pair is run exactly once and no averaging over repeated rollouts is needed. The Organizers' end-to-end re-run of the top 7 submitted policies at a fixed simulator commit therefore verifies rule compliance rather than estimating variance.

• No single baseline is best everywhere. On the public training scenarios, bang_bang sets the lower end of the performance range, while the upper end is set by pid_water or, in hot and humid weather, by fixed_setpoint — which the authors present as confirmation of the transfer challenge. The baselines do not enter the leaderboard metric.

• No empirical model results are reported. The paper states the shipped baseline returns for the public training scenarios exist, but does not report their numeric values; it reports no leaderboard standings or model performance figures.

• Ventilation floors prevent a degenerate "shut everything off" strategy. The simulator enforces a minimum static pressure for a running AHU, at most 70% recirculation, and dampers at least 10% open.

• Action and observation spaces are hybrid and building-dependent. For a building with K AHUs and V zones, the action dict has discrete_action of shape (K+1,) and continuous_action of shape (3K+2+2V,); continuous setpoints are normalized to [−1, 1]. The policy also receives building metadata at runtime, including the floor plan, AHU-to-zone assignment, action and observation indices, schedule, and normalization constants.

Methodology in Plain English

Rather than build a controller themselves, the authors design a contest that forces transfer.

First, they build a random building generator on top of an existing open-source HVAC simulator. Each generated building is a plausible office with its own floor count, footprint, room layout, occupancy pattern, envelope materials, and arrangement of air handlers and zones. Because generation is procedural, there is no fixed dataset to memorize — participants can sample as many training buildings as they want, and the Organizers can generate fresh evaluation buildings.

Second, they split by building configuration rather than by time. Training buildings and evaluation buildings come from the same generator and the same parameter ranges, but evaluation buildings are produced from seeds withheld from participants and never appear in any released artifact. The policy sees an evaluation building's floorplan, AHU divisions, and action spec only at runtime inside the evaluation sandbox, so it cannot precompute per-building heuristics.

Third, they define a submission interface. Contestants submit a Python class subclassing the evaluation gateway's policy base class, with a constructor that receives the building metadata and action spec once per building, and a get_action(obs) method that maps the current observation to an action dict. Submissions take the form of a Kaggle notebook; a submission contains no scenario list and no rollout loop, since the inference server owns the simulator and queries the policy once per step.

Fourth, they define scoring. Each scenario runs 30 days at a 5-minute timestep, every step is scored (T = 8640), and the score is the mean per-step reward's magnitude aggregated over 30 held-out scenarios. Four hand-written reference controllers — Bang-Bang, Fixed Setpoint, PID, and PID + Water — plus an "All Off" reference point that is not counted among the four, ship in the starter kit so participants have a published range to develop against. Contestants can develop locally against evaluate.py and score_leaderboard.py.

The competition runs in two phases: a development phase with a rotating, publicly-visible subset of evaluation pairs on the public leaderboard, and a final phase where all submissions are re-evaluated against the complete held-out set on the private leaderboard. Anti-cheating measures include never releasing the held-out 30 scenarios before the close, running policies in an isolated container with no network access, bounding per-rollout wall-clock at 10× the bang-bang baseline runtime, and scanning submitted code archives for hard-coded building IDs or sensor signatures.

Why This Matters

Impact on research. The paper positions building control as a concrete, high-stakes testbed for transfer learning and foundation models outside the usual vision, language, and robotics domains. It also poses an explicit open scientific question: whether domain-specific models will win out or whether generic foundation models will generalize best. If building foundation models succeed, the paper argues this is a model for how other real-world control problems could be framed as community benchmarks.

Real-world applications:

• Portfolio-wide building management. The task mirrors managing a diverse real-estate portfolio where per-building tuning is too costly and risky; a successful policy acts as a universal controller that adapts to unseen buildings using only floorplans and HVAC inventory.
• Lowering adoption costs. A transferable model removes the need for hard-to-calibrate per-building simulations, which the authors identify as a barrier to adopting optimal control.
• Emissions reduction at national scale. Buildings cause nearly 40% of US carbon emissions and commercial buildings alone account for 17% of total US carbon emissions; the authors estimate that a model unlocking even a 5% reduction in carbon emissions could reduce total US carbon emissions by roughly 1%.
• Post-contest live trials. Winning solutions will be evaluated on real buildings after the contest, at four physical partner sites, under strict institutional safety protocols so participants never directly control physical equipment.

Industry relevance. The prize pool is jointly sponsored by Kajima, Distyl, and KODE Labs (collectively the "Organizers"), and the author list spans Columbia, Harvard, Princeton, Mila, Syracuse, Google, Autodesk Research, and Kajima — a mix of academic labs, industry research groups, and construction/facilities firms. Top teams receive an invited podium talk at the NeurIPS 2026 Competition Track workshop, co-authorship on the consortium write-up (named-author status for teams contributing a methods description meeting the standard threshold, consortium-author status otherwise), and an invitation to the post-contest live-trial program.

Future Directions

• Do domain-specific or generic foundation models win? The authors explicitly frame this as an open question they hope the competition will help answer.
• Transfer from simulation to real buildings. The contest is run entirely in simulation, with successful solutions evaluated on real buildings only after the competition ends — leaving sim-to-real transfer unresolved within the paper.
• Whether a building foundation model is achievable at all. The central question the competition asks is whether a single controller can condition its decisions on the specific building it operates and transfer across many buildings.
• Community benchmark adoption. The authors aim to foster a well-adopted benchmark within the building science and ML communities, citing the AI City Challenge as a successful recent example in the adjacent ML-and-sustainability space, and noting that progress in other domains is bolstered when a community stands behind a benchmark.

Target Audience

This paper is aimed at machine learning researchers interested in transfer learning, reinforcement learning, and foundation models applied to real-world control; building scientists and HVAC practitioners; sustainability and climate-focused AI researchers; and competition participants ranging from RL specialists to newcomers, since the authors state that no domain knowledge is needed to participate. Industry readers evaluating scalable building-optimization approaches, and organizers of other real-world ML benchmarks, would also find the protocol and evaluation design relevant.

Schedule (as listed): Nov 13, 2026 — Round One deadline (NeurIPS poster; $2K in prizes); Dec 13, 2026 — Paris Workshop; Jan 15, 2027 — Final deadline ($38K in prizes). The $40,000 total prize pool is split across podium positions on the private leaderboard. Starter kit: https://github.com/smart-buildings-2026/starter-kit

Authors’ abstract

Buildings cause nearly 40\% of carbon emissions in the US. Of total building emissions, HVAC systems are responsible for approximately 40-60\% and account for 15\% of global energy consumption. Despite advances in artificial intelligence, most buildings still operate as inefficiently as they did decades ago. Building optimization, perhaps more so than any other optimization problem, has enormous potential for Machine Learning (ML) driven impact on climate change, in both the short and long term. One of the biggest obstacles for building optimization is the lack of policy transfer. Most building control systems, whether simulation or learning based, don't generalize: they only work well for the building they were developed for, but not for other buildings. To solve this problem, we propose a challenge for the development of building foundation models that easily adapt to new buildings with minimal configuration and thereby enable building optimization to scale across the globe. Participants will build control policies on a set of simulated buildings, and have their policies evaluated on an unseen set of buildings. The buildings vary in size, layout, weather, material properties, and HVAC system. Successful outcomes from this initiative promise a massive leap forward in global sustainability efforts and demonstrate the great potential of ML to combat climate change.