admin/lerobot
1
0
Fork 0
mirror of https://github.com/huggingface/lerobot.git synced 2026-05-22 03:59:42 +00:00
Code Issues Packages Projects Releases Wiki Activity

tune(annotations): VQA emission anchors a single frame (K 3 -> 1)

Browse Source 
Module 3 anchored each VQA emission tick to K=3 consecutive frames
(~0.1s at 30fps). The VLM grounds the answer — bbox/keypoint
coordinates especially — against the first frame's image, so copying it
onto frames 2-3 smears a stale label over a moving scene.

Default K=1: a VQA pair lands on exactly its emission frame, no
temporal smear. VQA frames get sparser; the WeightedEpisodeAwareSampler
(vqa_target_fraction) is the knob to compensate.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Pepijn
2026-05-18 17:24:36 +02:00
parent 0f5f0e4091
commit 474c5478d9
1 changed files with 8 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/lerobot/annotations/steerable_pipeline/config.py
+8 -1
View File
@@ -114,7 +114,14 @@ class Module3Config:
enabled: bool = True enabled: bool = True
vqa_emission_hz: float = 1.0 vqa_emission_hz: float = 1.0
K: int = 3 K: int = 1
"""How many *consecutive* frames each emission tick anchors a VQA pair
to. The VLM grounds its answer (bbox / keypoint coordinates, count, …)
against the *first* anchored frame's image, so anchoring K>1 frames
copies that same answer onto later frames where the scene has already
moved — stale labels. Default ``1``: a VQA pair lands on exactly its
emission frame, no temporal smear. Raise it only to trade label
precision for more (noisier) VQA frames."""
question_types: tuple[str, ...] = ("bbox", "keypoint", "count", "attribute", "spatial") question_types: tuple[str, ...] = ("bbox", "keypoint", "count", "attribute", "spatial")