fix(rtc): bootstrap trained inference after cold start

This commit is contained in:
Pepijn
2026-07-17 15:45:53 +02:00
parent 962d754658
commit a0ae756221
2 changed files with 70 additions and 17 deletions
+26 -1
View File
@@ -110,12 +110,37 @@ def test_trained_rtc_retries_chunk_when_measured_delay_exceeds_conditioning():
)
assert _trained_rtc_chunk_can_merge(
conditioned_delay=2,
measured_delay=3,
measured_delay=5,
training_max_delay=4,
has_previous_actions=False,
)
def test_trained_rtc_bootstraps_first_overlap_with_checkpoint_capacity():
from lerobot.rollout.inference.rtc import _estimate_rtc_delay
assert (
_estimate_rtc_delay(
latency=0,
time_per_step=1 / 30,
mode="trained",
training_max_delay=10,
has_previous_actions=False,
)
== 0
)
assert (
_estimate_rtc_delay(
latency=0,
time_per_step=1 / 30,
mode="trained",
training_max_delay=10,
has_previous_actions=True,
)
== 10
)
def test_trained_rtc_rejects_measured_delay_above_checkpoint_support():
from lerobot.rollout.inference.rtc import (
_trained_rtc_chunk_can_merge,