Demonstration timing: telemetry to latency-corrected action label
Labels are constructed on the same physical decision-time grid used online; lead and aggregation change the label, while offset is an open-loop diagnostic.
Recorded frame_start stream · irregular telemetry timestamps
Decision-grid sample construction
Evaluation paths
frame_start
held
new sample
false
true
diagnose phase
candidate offset
feedback policy
Telemetry F_j @ τ_j
position · velocity ·
race timer
Control U_j
held from τ_j
Telemetry F_{j+1}
next recorded
timestamp
Control U_{j+1}
held from τ_{j+1}
… later frames
monotonic race-time
axis
Decision tick t_k
next tick ≥ t_k +
decision_interval_ms
Observation x_k
feature pipeline on
selected telemetry
Shifted label time
τlabel = t_k +
action_lead_ms
Aggregate
controls?
config flag
Point label
first recorded action
at or after τlabel
Window label
integrate window
then quantize once
BC training sample
x_k + discrete label
online cadence
Open-loop replay
action selected by
current race timer
--action-offset-ms
sweep replay timestamps
to diagnose latency
Trajectory tracking
state-based feedback
corrects lateral error
Closed-loop benchmark
finish · pace ·
interventions · failures
Aggregation requires control_alignment=frame_start and integrates [t_k +
lead, t_{k+1} + lead). Gas/steering use duration-weighted means; brake taps
use overlap duration before quantization.
--action-offset-ms shifts only open-loop replay timestamps for diagnosis. It
does not rewrite the demonstration or replace closed-loop bc-benchmark.
Download editable diagram