beet_ml

beet_ml brings machine learning into the action model. Like beet_spatial it builds on beet_action, so a model invocation is an action that slots into a behavior tree alongside everything else.

It currently spans two domains:

• Language: the Bert sentence-embedding asset, running on burn with selectable wgpu, ndarray or cuda backends, paired with a Sentence action that picks the closest match to a user phrase. This is enough to route natural language to behavior without a cloud round trip.
• Reinforcement learning: a FrozenLake environment and Q-learning agents ported from OpenAI Gym, runnable headless for training or in realtime to watch them learn.

Add BeetMlPlugins to register the assets, actions and tick schedule.