Estimates the value of a trajectory as per eqn 4 of the FOWM paper. Args: z: (batch, latent_dim) tensor of initial latent states.

Flatten a nested dictionary structure by collapsing nested keys into one key with a separator. For example: ``` >>> dct = {"a": {"b": 1,

This function expects `batch` to have: { "observation.state": (B, n_obs_steps, state_dim) "observation.image

Returns all data for a given data key as a Tensor.

Get a module's parameter dtype by checking one of its parameters. Note: assumes that all parameters have the same dtype.

Common shapes T,D T,N,D T,H,W,C T,N,H,W,C

Get the policy's class and config class given a name (matching the policy class' `name` attribute).

These einops patterns will be used to aggregate batches and compute statistics. Note: We assume the images are in channel first format

Predict the next state's latent representation given a current latent and action. Args: z: (*, latent_dim) tensor for the current

episode_data_index contains the range of indices for each episode Example: ```python from_id = episode_data_index["from"][episode_id].ite

Note: When using data workers (e.g. DataLoader with num_workers>0), do not call this function in the main process (e.g. by using a second Dataload

hf_dataset contains all the observations, states, actions, rewards, etc.

info contains useful information regarding the dataset that are not stored elsewhere Example: ```python print("frame per second used to c

Given the last checkpoint in the logging directory, load the optimizer state, scheduler state, and random state, and return the globa

Given a current item in the dataset containing a timestamp (e.g. 0.6 seconds), and a list of time differences of some modalities (e.g. delta_

stats contains the statistics per modality computed over the full dataset, such as max, min, mean, std Example: ```python normalized_acti

for given ground truth action values (target), and prediction (pred) this function calculates the overall loss. predicted_action: pr

Plan sequence of actions using TD-MPC inference. Args: z: (batch, latent_dim,) tensor for the initial state. Returns:

Creates and pushes a LeRobotDataset Card with appropriate tags to easily find it on the hub.

Expect all meta data files to be all stored in a single "meta_data" directory. On the hugging face repositery, they will be uploaded in a "meta_da

Expect mp4 files to be all stored in a single "videos" directory. On the hugging face repositery, they will be uploaded in a "videos" directory at

Register codecs in this module with numcodecs.

Clear observation and action queues. Should be called on `env.reset()`

Clear observation and action queues. Should be called on `env.reset()` queues are populated during rollout of the policy, they contai

Clear observation and action queues. Clear previous means for warm starting of MPPI/CEM. Should be called on `env.reset()`

This should be called whenever the environment is reset.

Resets the online computation variables.

Reset the `episode_index` of the provided HuggingFace Dataset. `episode_data_index` (and related functionality such as `load_previous_and_future_

Inverse of `apply_calibration`.

Run a batched policy rollout once through a batch of environments. Note that all environments in the batch are run until the last environment is

Checkpoint the model weights and the training state.

Save the weights of the Policy model using PyTorchModelHubMixin. The weights are saved in a folder called "pretrained_model" under the checkp

Checkpoint the global training_step, optimizer state, scheduler state, and random state. All of these are saved as "training_state.pth" under

Set delta_timestamps converting the values to numpy arrays. The conversion is for an optimization in the __getitem__. The loop is much slower

Set the freedrive mode of the robot. Args: enable (bool): True to enable freedrive mode, False to disable it.