Imitation learning lifecycle and controlled RL handoff
Behavior cloning is the supervised algorithm; imitation_learning owns the wider data, recovery, evaluation and warm-start workflow.
Trusted demonstration data
Behavior-cloning lifecycle
Closed-loop gate and RL
experiment
training set
held-out laps
epochs
candidate
metrics
pass
fail / intervene
record-demo
whole laps ·
outlier filtering
DAgger recovery
student states ·
expert labels
Dataset quality
complete laps ·
timing
Dataset manifest
fingerprint ·
provenance
Lap-level split
train / validation;
no leakage
Safe augmentation
horizontal flip;
preserve unknowns
BehaviorCloningModel
encoder · temporal ·
actor head
Offline learner
contiguous batches ·
AMP · scaler
Validation
aggregate
loss numerator /
sample count
BC checkpoint
exact trainer state
Exact resume
same split, order
and next update
Closed-loop eval
finish · pace ·
intervention rate
Promotion gate
named criteria, never
implicit
Warm-start RL
encoder + temporal;
new value head
Package boundary: trackmaniarl.trackmania.imitation_learning. Public classes include LidarBehaviorCloningModelFactory and
BehaviorCloningLearner.
Warm-start is partial loading with an explicit match report; it is not a resume and it does not transfer optimizer state or episodic
hidden state.
Download editable diagram