policy

policy

PolicyConfig dataclass

PolicyConfig(_target_='versatil.models.policy.Policy', encoding_pipeline=MISSING, algorithm=MISSING, decoder=MISSING, observation_space='${task.observation_space}', action_space='${task.action_space}', prediction_horizon='${task.prediction_horizon}', observation_horizon='${task.observation_horizon}', device='${experiment.device}', loss=MISSING, metadata_passthrough=dict())

Hydra config for constructing a policy from encoding, algorithm, decoder, and loss configs.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
encoding_pipeline	EncodingPipelineConfig	Observation encoding pipeline.
algorithm	Any	Decoding algorithm (diffusion, flow matching, etc.).
decoder	Any	Action decoder architecture.
observation_space	ObservationSpaceConfig	Observation space configuration.
action_space	ActionSpaceConfig	Action space configuration.
prediction_horizon	int	Number of future actions to predict.
observation_horizon	int	Number of past observations to condition on.
device	str	Device to run on.
loss	CompositeLossConfig	Loss module for training.
metadata_passthrough	dict[str, dict[str, str]]	Mapping from source dictionaries to metadata keys for logging/visualization.