AgriManager
A multi-simulator LLM-RL framework for agricultural management.
AgriManager integrates large language models with agricultural reinforcement-learning environments. It unifies the DSSAT, WOFOST, and Cycles crop simulators behind a single prompt-based text interface, so one reasoning policy can be trained jointly across simulators, objectives, and interfaces — using GRPO via VeRL.
What is AgriManager#
A crop-management policy faces heterogeneity of two kinds. Distributional shift keeps the interface fixed while weather, crop, or prices change. Schema shift changes the interface itself — different farms expose different observed variables, actions, or objectives.
A fixed-interface neural policy cannot run once a shift changes its input layout or action encoding. A prompt-conditioned LLM policy re-reads the task schema from the prompt of each episode, so a single trained policy can operate across interfaces that would otherwise each need their own architecture. AgriManager is the framework — and benchmark — for studying exactly this.
Highlights#
- One text interface over three simulators — DSSAT, WOFOST, and Cycles, via their gym wrappers.
- Schema declared in the prompt — adapt to a new observation set, action menu, or objective with no architectural change.
- Joint multi-objective training — GRPO on VeRL, plus a framework-native neural-network PPO baseline for controlled comparison.
- A three-tier generalization benchmark — eight settings from Within Schema to Compositional Schema shift, scored on a frozen held-out test suite with three training seeds per setting.
- Pluggable environments — add an env by dropping a four-file package into
agrimanager/env/; there is no registry.