Rollout

Running a policy in an environment from start to finish to produce one complete interaction trajectory.

A rollout means placing a policy in an environment, whether simulated or real, and running the “observe, act, environment changes” loop from the initial state until the episode succeeds, fails, or times out, producing one complete trajectory. OpenAI's Spinning Up tutorial notes that a trajectory is also commonly called an episode or a rollout. Common usages: in reinforcement learning, rollouts are used to collect the data that updates a policy; in evaluation, “50 rollouts per task” means running 50 episodes and computing the success rate from them; and in model-based reinforcement learning, taking several imagined steps forward inside a world model is also called a rollout. Reinforcement learning for large language models has carried the same term over to mean sampling several responses for the same prompt.

ExampleThe RoboChallenge Table30 benchmark runs 10 rollouts per task, then scores each model by success rate and progress score.

Related

Trajectory · Episode · Success Rate · Closed-Loop Evaluation · Learning in Imagination · On-Policy

Sources

OpenAI Spinning Up: Key Concepts in RL
RoboChallenge 论文 HTML 全文 (arXiv 2510.17950) (Chinese)