Struct RolloutBatch 

pub struct RolloutBatch {
    pub observations: Vec<f32>,
    pub actions: Vec<i64>,
    pub old_log_probs: Vec<f32>,
    pub old_values: Vec<f32>,
    pub advantages: Vec<f32>,
    pub returns: Vec<f32>,
}

Batch of rollout data for training

Contains flattened tensors suitable for neural network training. All arrays have shape [batch_size].

Fields

observations: Vec<f32>

Flattened observations [batch_size, obs_dim]

actions: Vec<i64>

Actions taken [batch_size]

old_log_probs: Vec<f32>

Old log probabilities [batch_size]

old_values: Vec<f32>

Old value estimates [batch_size]

advantages: Vec<f32>

Computed advantages [batch_size]

returns: Vec<f32>

Computed returns [batch_size]

Implementations

impl RolloutBatch

pub fn from_buffer(buffer: &RolloutBuffer) -> Self

Create a new batch from rollout buffer

Iterates the full [num_steps, num_envs] capacity. If the buffer was only partially filled, the unwritten tail surfaces as zero-initialized rows. Use Self::from_buffer_partial (or super::RolloutBuffer::get_filled_batch) when the caller knows the fill count.

pub fn from_buffer_partial(buffer: &RolloutBuffer, valid_steps: usize) -> Self

Create a new batch from the first valid_steps rows of the buffer.

Rows in valid_steps..num_steps (the unfilled tail of a partial rollout) are skipped, preventing zero-padded rows from contaminating PPO gradients.

Panics

Panics if valid_steps > buffer.num_steps.

pub fn len(&self) -> usize

Get batch size

pub fn is_empty(&self) -> bool

Check if batch is empty

pub fn obs_shape(&self) -> (usize, usize)

Get the observation shape as (batch_size, obs_dim)

pub fn to_burn_tensors<B: Backend>( &self, device: &B::Device, ) -> RolloutBurnTensors<B>

Construct the full set of training tensors from this batch in a Construct the full set of training tensors as Burn tensors on device.

Returns a named RolloutBurnTensors bundle so trainers can pattern-match named fields rather than positional tuple elements.

Shapes (all on device):

• observations: [batch, obs_dim], f32
• actions: [batch], i64 (Burn Int kind)
• old_log_probs: [batch], f32
• old_values: [batch], f32
• advantages: [batch], f32
• returns: [batch], f32

Empty batches still produce well-formed [0, obs_dim] / [0] tensors. The obs_dim is derived from Self::obs_shape (which returns 0 for an empty batch), so the observation tensor for the empty case is shaped [0, 0].

Trait Implementations

impl Clone for RolloutBatch

fn clone(&self) -> RolloutBatch

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for RolloutBatch

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.