An open-source toolkit for using Unity environments to train and evaluate agents. Builders define observations, actions, and training signals for a specific sim
Last reviewed: 2026-10-03
Unity ML-Agents connects Unity environments with learning algorithms and training workflows. It supports reinforcement and imitation learning, including single-agent and multi-agent scenarios. The toolkit provides training infrastructure; the builder defines the environment and learning problem.
Start with a small scene and document the observations, available actions, rewards, and episode termination. Compare the trained policy with a simple scripted agent on unseen scenarios. Look for reward exploitation and behavior that works only in the training scene before increasing complexity.
No. You need to define the environment, observations, actions, training signals, and evaluation.
Watch complete episodes and compare task success with the reward score. Test altered scenarios where shortcuts from the training environment no longer work.