LeRobot PR

LeRobot PR #3179 / video timestamp fix

English

Pull request by joonmo-han (weedmo) fixing lerobot-issue-3177:

• Title: fix(datasets): use actual video duration for timestamps instead of float accumulation
• URL: https://github.com/huggingface/lerobot/pull/3179
• Author: weedmo
• Status (as of 2026-09-26): open, not merged
• Linked issue: https://github.com/huggingface/lerobot/issues/3177

Change summary (from author comment on the issue):

• lerobot_dataset.py: read latest_duration_in_s from the actual video via get_video_duration_in_s() instead of parquet metadata; recompute episode duration from total duration after concatenation
• aggregate.py: same pattern for dataset aggregation
• Regression test: 50 episodes with varying frame counts (7–17), assert from_timestamp == actual video PTS at boundaries

Outcome so far: Issue closed as stale without merge. Collaborator reported suggested tests already pass on main. Contribution record: authored bug report + proposed fix PR (pending review/merge).

한국어

lerobot-issue-3177를 고치기 위해 joonmo-han (weedmo)가 연 PR입니다.

• 제목: fix(datasets): use actual video duration for timestamps instead of float accumulation
• URL: https://github.com/huggingface/lerobot/pull/3179
• 작성자: weedmo
• 상태 (2026-09-26 기준): open, 미머지
• 연결 이슈: https://github.com/huggingface/lerobot/issues/3177

변경 요약 (이슈 댓글 기준):

• lerobot_dataset.py: parquet 메타 대신 실제 비디오에서 get_video_duration_in_s()로 duration 읽고, concat 후 총 length로 에피소드 duration 재계산
• aggregate.py: 동일 패턴
• 회귀 테스트: 프레임 수 7–17인 에피소드 50개, 경계에서 from_timestamp가 실제 PTS와 일치하는지 검증

현재까지 결과: 이슈는 stale로 닫힘, 미머지. Collaborator는 제안 테스트가 main에서도 통과한다고 함. 기여 기록: 버그 리포트 작성 + 수정 PR 제안 (리뷰/머지 대기).