Other Loss Modules

Additional loss modules for specialized algorithms.

ACTLoss(*args, **kwargs)	Loss module for Action Chunking with Transformers (ACT).
BCLoss(*args, **kwargs)	Behavior Cloning Loss Module.
DiffusionBCLoss(*args, **kwargs)	Behavioural Cloning loss for diffusion-based policies.
GAILLoss(*args, **kwargs)	TorchRL implementation of the Generative Adversarial Imitation Learning (GAIL) loss.
DTLoss(*args, **kwargs)	TorchRL implementation of the Online Decision Transformer loss.
OnlineDTLoss(*args, **kwargs)	TorchRL implementation of the Online Decision Transformer loss.
DreamerActorLoss(*args, **kwargs)	Dreamer Actor Loss.
DreamerModelLoss(*args, **kwargs)	Dreamer Model Loss.
DreamerValueLoss(*args, **kwargs)	Dreamer Value Loss.
ExponentialQuadraticCost(*args, **kwargs)	Computes the expected saturating cost for a Gaussian-distributed state.

DreamerV3

Loss modules for DreamerV3 (Mastering Diverse Domains in World Models, Hafner et al. 2023). Key differences from V1: discrete categorical latent state, KL balancing, symlog transforms, and two-hot value distributions.

DreamerV3ActorLoss(*args, **kwargs)	DreamerV3 Actor Loss.
DreamerV3ModelLoss(*args, **kwargs)	DreamerV3 World Model Loss.
DreamerV3ValueLoss(*args, **kwargs)	DreamerV3 Value Loss.

DreamerV3 Utilities

symlog(x)	Symmetric logarithm: sign(x) * log(|x| + 1).
symexp(x)	Symmetric exponential: sign(x) * (exp(|x|) - 1).
two_hot_encode(x, bins)	Encode a scalar tensor as a two-hot distribution over bins.
two_hot_decode(logits, bins)	Decode a distribution over bins to a scalar expectation.