TRACKMANIARL / SYSTEM GUIDE / 05
Align demonstrations with action time
Construct labels on the same physical decision grid used by the online policy.
01 Align the sample in physical time
02a Point label
02b Window label
Recorded telemetry
frame_start timestamps
Held controls
Decision tick t_k
Feature pipeline
Decision interval
Shifted label time
Δlead = action_lead_ms
First action at / after t_label
Use when control aggregation is disabled.
Aggregate, then quantize
Integrate shifted decision windows.
Weight gas / steering by duration.
Behavior-cloning sample
Observation x_k + one discrete action label, at the online decision cadence.
WINDOW CONTRACT
Use control_alignment=frame_start. Brake taps
use overlap duration before quantization.
DIAGNOSTIC ONLY
--action-offset-ms shifts open-loop replay.
Validate latency choices in closed-loop
benchmarks.
Download editable diagram