RL SMPC Blends RL and Stochastic MPC for Greenhouse Lettuce

What changed

Bart van Laatum and coauthors published a peer reviewed study in Control Engineering Practice, volume 169, article 106787, in 2026. The DOI is 10.1016/j.conengprac.2026.106787. The title covers model predictive control with reinforcement learning for greenhouse crops under uncertain parameters.

The companion GitHub repo BartvLaatum/RL-SMPC implements RL-SMPC on a lettuce greenhouse model. A trained RL policy supplies nominal control trajectories. A random scenario MPC layer then optimizes corrections when weather and crop parameters drift.

The workflow trains RL agents and fits value functions for end of horizon costs. It then compares RL, MPC, SMPC, and RL-SMPC across eight horizons and ten random seeds. Scripts live under run_scripts with a full evaluation shell that can run for days on strong hardware.

Why it matters

Greenhouse operators must balance heat, CO2, screens, and profit when weather forecasts are wrong. Pure RL can adapt but may break hard climate limits. Pure MPC can be slow when the crop model drifts. RL-SMPC blends a learned policy with a solver step. That fits high tech lettuce lines that already use model based control.

Published code lowers the bar for other crop models to test the same hybrid loop without rebuilding solvers from scratch.

Who is affected

Agri tech R&D teams building climate controllers for leafy greens should read the paper and repo scripts. Greenhouse automation vendors comparing RL offerings to MPC baselines can reuse the evaluation shell. Academic groups teaching greenhouse control have a citable lettuce case with open scripts.

Grower co ops that rely on vendor black boxes may still care because the paper names economic reward terms tied to marketable yield. That language helps procurement ask better questions.

What to do next

Clone RL-SMPC and read the README workflow for training, value function fitting, and ./run_scripts/execute_all.sh. Start with a short horizon and one seed before you launch the full multi day sweep described in the repo.

Map the economic reward and constraint set to your own greenhouse digital twin. If your model differs from the lettuce environment in the repo, treat the paper as a method reference rather than a plug in controller. Save raw CSV outputs from the uncertainty comparison folder for audit.

What to watch

Watch for follow on releases from the same authors on tomato or other crops in GreenLight Gym repos. Watch whether commercial greenhouse software cites RL-SMPC in 2026 conference talks. Watch the journal page for corrigenda tied to article 106787.

Watch local energy prices in your region because the reward function weights profit against heat and light use. A policy tuned in the paper may need retuning when tariffs shift.

The repo cites Wageningen affiliations in the README contact line. Teams outside the Netherlands should still validate weather files and greenhouse geometry before they trust default parameters. The lettuce model builds on prior GreenLight style dynamics, so tomato first operators should not copy heat setpoints without review.

If you lack GPU farms for RL training, start with the MPC and SMPC baselines in the same script bundle. They provide lower compute anchors that help you judge whether RL-SMPC gains justify training cost on your crop mix.

The paper authors list an email contact in the repo README for research questions. Vendor teams should still run their own shadow mode trials before they push any new setpoints to live vents or lamps. Field pilots need backup manual rules when the optimizer hits a constraint wall during a heat wave.

The bibtex block in the repo matches the journal citation. That helps lit reviews stay aligned with the code drop. Students can try other leaf crops only after they rebuild the reward model with local trial data.

Sensor vendors can use the open scripts as a benchmark hook. They should still prove their humidity probes match the sim before they claim parity with the paper curves.

Photo: Tom Fisk / Pexels

Sources

• Primary — van Laatum et al., Control Engineering Practice article 106787 (2026). Journal version of RL-SMPC for greenhouse lettuce under uncertainty.
• Primary — Bart van Laatum, RL-SMPC GitHub repository (2026). Code, workflow, and bibtex citation matching the paper.
• Secondary — Elsevier, Control Engineering Practice journal home (2026). Publisher context for volume 169.