algorithm

algorithm

Action-decoding algorithm configurations.

DecodingAlgorithmConfig dataclass

DecodingAlgorithmConfig(_target_=MISSING)

Base algorithm configuration.

Note: For variational algorithms, use VariationalAlgorithmConfig instead of setting latent_encoder on individual algorithms.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.

BehavioralCloningConfig dataclass

BehavioralCloningConfig(_target_='versatil.models.decoding.algorithm.behavior_cloning.BehavioralCloning')

Bases: DecodingAlgorithmConfig

Behavioral Cloning (direct supervised prediction) algorithm configuration.

This is a pure, deterministic algorithm. For multi-modal action prediction, use VariationalAlgorithmConfig with BehavioralCloningConfig as the base_algorithm.

DiffusionConfig dataclass

DiffusionConfig(_target_='versatil.models.decoding.algorithm.diffusion.Diffusion', scheduler_type=value, num_train_timesteps=100, num_inference_steps=10, beta_start=0.0001, beta_end=0.02, beta_schedule=value, prediction_type=value, scheduler_variance_type=value, clip_sample=True, set_alpha_to_one=True, steps_offset=0)

Bases: DecodingAlgorithmConfig

Diffusion algorithm configuration.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
scheduler_type	str	Type of diffusion scheduler ("ddpm" or "ddim").
num_train_timesteps	int	Number of diffusion steps during training.
num_inference_steps	int	Number of denoising steps during inference.
beta_start	float	Starting value of noise schedule.
beta_end	float	Ending value of noise schedule.
beta_schedule	str	Noise schedule type ("linear", "squaredcos_cap_v2", etc.).
prediction_type	str	What the network predicts ("epsilon" for noise, "sample" for clean actions).
scheduler_variance_type	str	Variance type for DDPM scheduler.
clip_sample	bool	Whether to clip samples to [-1, 1] during inference.
set_alpha_to_one	bool	Whether to set final alpha to 1.
steps_offset	int	Offset for timestep calculation.

FlowMatchingConfig dataclass

FlowMatchingConfig(_target_='versatil.models.decoding.algorithm.flow_matching.FlowMatching', sigma=0.0, ode_solver=value, num_inference_steps=10, timestep_sampler=BETA.value, logit_mean=0.0, logit_std=1.0, beta_alpha=1.5, beta_beta=1.0, max_timestep=0.999, reverse_flow_convention=False)

Bases: DecodingAlgorithmConfig

Flow Matching algorithm configuration.

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
sigma	float	Noise level for conditional flow matching (0 = deterministic OT).
ode_solver	str	ODE solver to use ("euler", "heun", or "rk4").
num_inference_steps	int	Number of integration steps during inference.
timestep_sampler	str	Timestep sampling strategy.
logit_mean	float	Mean for logit-normal (shifts mode; 0 centers at t=0.5).
logit_std	float	Std for logit-normal (smaller = more concentrated).
beta_alpha	float	First shape parameter for Beta distribution (pi0 uses 1.5).
beta_beta	float	Second shape parameter for Beta distribution (pi0 uses 1.0).
max_timestep	float	Upper bound s for Beta sampling (pi0 uses 0.999).
reverse_flow_convention	bool	Reverse the time/velocity convention during inference. When True, the inference loop remaps t to (1-t) and negates the predicted velocity.

VariationalAlgorithmConfig dataclass

VariationalAlgorithmConfig(_target_='versatil.models.decoding.algorithm.variational.VariationalAlgorithm', base_algorithm=MISSING, posterior_encoder=MISSING, prior=MISSING, sampling_from_prior_probability=0.25, posterior_decoder_noise_std=0.0)

Bases: DecodingAlgorithmConfig

Compositional variational inference wrapper configuration.

Wraps any base algorithm with variational latent encoding for multi-modal action prediction. This replaces the need for algorithm-specific variational implementations.

Examples:

Behavioral Cloning with VAE + Gaussian prior

VariationalAlgorithmConfig( base_algorithm=BehavioralCloningConfig(), posterior_encoder=VAETransformerEncoderConfig(...), prior=None # Auto-creates GaussianPrior )

Flow Matching with VAE + Diffusion prior (replaces VariationalFlowMatching)

VariationalAlgorithmConfig( base_algorithm=FlowMatchingConfig(sigma=0.0, num_inference_steps=10), posterior_encoder=VAETransformerEncoderConfig(...), prior=DiTPriorConfig(...) )

NEW: Diffusion with VAE + learned prior

VariationalAlgorithmConfig( base_algorithm=DiffusionConfig(...), posterior_encoder=VAETransformerEncoderConfig(...), prior=DiTPriorConfig(...) )

Attributes:

Name	Type	Description
_target_	str	Import path instantiated by Hydra.
base_algorithm	DecodingAlgorithmConfig	The base decoding algorithm (BC, FlowMatching, Diffusion, etc.)
posterior_encoder	PosteriorLatentEncoderConfig	Latent encoder for posterior q(z|a,s) (typically VAE)
prior	PriorLatentEncoderConfig	Latent prior for p(z|s). If None, auto-creates GaussianPrior.