Training on Arena

Arena is AgileRL’s managed RLOps platform for running reinforcement learning on cloud infrastructure. It is built for the same evolutionary training pipeline you use locally with LocalTrainer, but runs jobs on Arena’s managed compute so you can train and scale your workloads efficiently without the headache of setting up a pipeline on a cloud provider or managing a cluster. Arena is provider-agnostic and gives users the flexibility to train on a variety of compute resources.

The Arena client is how you interact with the platform from your machine: upload and validate custom environments, submit training jobs on scalable compute, and deploy trained agents for inference.

See also

Arena Client for the full CLI and Python SDK reference (commands, manifests, and API details).

Trainers for how LocalTrainer and ArenaTrainer relate to local and cloud training.

Tutorials

• PPO on a Custom Gym Environment: Upload and validate a Merge air traffic environment on Arena, train PPO with evolutionary HPO, and deploy the trained agent for inference.

• GRPO on GSM8K: Register the GSM8K math dataset from Hugging Face, write a reward function, fine-tune Qwen2.5-0.5B-Instruct with GRPO, and deploy the model for inference.