admin/lerobot
1
0
Fork 0
mirror of https://github.com/huggingface/lerobot.git synced 2026-07-24 18:26:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

annotate: make full-episode subtask coverage unconditional

Browse Source 
Remove the subtask_full_coverage config flag. Stitching subtask spans
into a contiguous full-episode cover is now always applied in
_generate_subtasks — a sparse / gap-ridden subtask timeline is never
desirable for conditioning, so there's no reason to make it optional.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Pepijn
2026-06-02 15:36:23 +02:00
parent 799d0e3bcc
commit 79f9a84407
2 changed files with 11 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/lerobot/annotations/steerable_pipeline/config.py
+6 -9
View File
@@ -73,15 +73,12 @@ class PlanConfig:
# nothing). +1 VLM call/ep. # nothing). +1 VLM call/ep.
subtask_verify: bool = True subtask_verify: bool = True
# ``subtask_full_coverage``: deterministic post-step (no VLM call) # NOTE: subtask spans are ALWAYS stitched into a contiguous
# that stitches the surviving subtask spans into a contiguous cover # full-episode cover (first subtask pulled back to t0, gaps closed,
# of the whole episode — first subtask pulled back to t0, each span's # last span extended to t_last) as a deterministic post-step in
# end snapped to the next span's start, last span extended to t_last. # ``_generate_subtasks._stitch_full_coverage``. This is not
# Without it the verify pass (which prunes spans) can leave the # configurable — a sparse / gap-ridden subtask timeline is never
# subtask timeline starting late or full of gaps, so frames fall # desirable for conditioning, so it is unconditional.
# through with no active subtask. On by default; disable only if a
# downstream consumer genuinely wants sparse (non-tiling) subtasks.
subtask_full_coverage: bool = True
# When True (and backend supports it, e.g. ``openai``), the ``plan`` # When True (and backend supports it, e.g. ``openai``), the ``plan``
# module sends a ``video_url`` block pointing at a per-episode mp4 # module sends a ``video_url`` block pointing at a per-episode mp4
src/lerobot/annotations/steerable_pipeline/modules/plan_subtasks_memory.py
+4 -3
View File
@@ -578,9 +578,10 @@ class PlanSubtasksMemoryModule:
# The VLM (especially after the verify pass prunes spans) can # The VLM (especially after the verify pass prunes spans) can
# leave the first subtask starting after t0 or leave gaps between # leave the first subtask starting after t0 or leave gaps between
# spans, so the subtask timeline no longer tiles the whole # spans, so the subtask timeline no longer tiles the whole
# episode and frames fall through with no active subtask. Stitch # episode and frames fall through with no active subtask. Always
# the surviving spans into a contiguous cover of [t0, t_last]. # stitch the surviving spans into a contiguous cover of
if getattr(self.config, "subtask_full_coverage", True): # [t0, t_last] — there is no scenario where a sparse, gap-ridden
# subtask timeline is desirable for conditioning.
cleaned = self._stitch_full_coverage(cleaned, record) cleaned = self._stitch_full_coverage(cleaned, record)
return cleaned return cleaned