Expand description
Fixed dataset of expert demonstrations for Behavioral Cloning.
Demonstrations is a small, flat supervised dataset of discrete-action
demonstrations: a contiguous block of observations paired with the expert
action taken at each. It is intentionally NOT a reuse of
crate::buffer::rollout::RolloutBuffer or
crate::buffer::replay::ReplayBuffer, which carry rewards, dones,
advantages, and priorities that supervised imitation has no use for (BC
epic #161, decision 2).
Seeded minibatching reuses PPO’s
generate_minibatch_indices_with_rng
so the BC trainer (#167) can produce bit-reproducible shuffles from a
BcConfig seed.
This is the dataset half of PR A of the BC decomposition (#164). A
from_file constructor can be layered on later without disturbing the
in-memory API.
Structs§
- Demonstrations
- A fixed dataset of expert
(observation, action)pairs for supervised imitation learning.