Struct RecurrentRolloutBatch 

pub struct RecurrentRolloutBatch<B: Backend> {
    pub obs_seq: Tensor<B, 3>,
    pub actions: Tensor<B, 2, Int>,
    pub episode_starts: Tensor<B, 2>,
    pub old_log_probs: Tensor<B, 2>,
    pub old_values: Tensor<B, 2>,
    pub advantages: Tensor<B, 2>,
    pub returns: Tensor<B, 2>,
}

A rank-3 batch of recurrent rollout data, ready to feed LstmBurnPolicy::evaluate_sequences with no shape adapters.

Every field is per-(env, step): obs_seq is rank-3 [N_env, T, obs_dim], the rest are rank-2 [N_env, T]. The recurrent PPO surrogate needs these per-step quantities, not a flattened rank-1 batch. initial_state is intentionally absent — the training forward always starts from a zeroed (h, c) and relies on episode_starts for in-sequence resets (design note Q2, Strategy A).

Fields

obs_seq: Tensor<B, 3>

Observations, [N_env, T, obs_dim] — feeds obs_seq.

actions: Tensor<B, 2, Int>

Discrete actions, [N_env, T] — feeds actions.

episode_starts: Tensor<B, 2>

Episode-start (state-reset) mask, [N_env, T]: 1.0 where obs[t] is the first step of a new episode, else 0.0. This is the done-flag stream shifted one step later — terminated[t-1] || truncated[t-1] for t >= 1, and the cross-iteration carry-in flag at t == 0. Feeds episode_starts, the hidden-state reset mask consumed by evaluate_sequences; it is distinct from (and one step ahead of) the terminated-only GAE flag.

old_log_probs: Tensor<B, 2>

Behavior-policy log-probs [N_env, T] for the PPO ratio.

old_values: Tensor<B, 2>

Behavior-policy value estimates V(s_t), [N_env, T].

advantages: Tensor<B, 2>

GAE advantages [N_env, T] (terminated-only bootstrap, from GAE).

returns: Tensor<B, 2>

Value-function targets [N_env, T] (advantages + values).

Implementations

impl<B: Backend> RecurrentRolloutBatch<B>

pub fn num_envs(&self) -> usize

Number of env-trajectories (N_env) in the batch.

pub fn seq_len(&self) -> usize

Sequence length (T) of each trajectory.

Trait Implementations

impl<B: Debug + Backend> Debug for RecurrentRolloutBatch<B>

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

Formats the value using the given formatter.