Get the value predictions from the model at each timestep.

Creates an actor. An actor is a `ConfigDict` with a positions `pos` and a direction `dir`. The position is an array with two elements, the he

Get the integer actions from the model at each timestep.

Generate a [num_steps x batch_size] array of generalized advantages using GAE.

Run the model for one timestep and return a dict of outputs. Args: states: a Tensor of previous states.

Copy this recording, but flip it left-to-right.

Reduces the 6 channels of `obs` to 3 RGB. Args: obs: the observation as a numpy array. Returns: An RGB image in the form of a numpy arra

Returns result of applying `self.operation` to a contiguous subsequence of the array. self.operation(arr[start], operation(arr[sta

Returns result of applying `self.operation` to a contiguous subsequence of the array. self.operation(arr[start], operation(arr[sta

Reset all the environments and return an array of observations, or a tuple of observation arrays. If step_async is still doin

Run the model on the rollout and create a new rollout with the filled-in model_outs. This may be more efficient than using s

Sample a batch of experiences. Parameters ---------- batch_size: int How many transitions to sample. Retur

Sample recordings such that recordings are weighted in proportion to their number of frames.

Truncate the recordings so that they never exceed a given floor. This can be used, for example, to train an agent to solve the begin