Fix Multi File Aggregate Datasets

[nicholas-maselli]

There is a bug in the src\lerobot\datasets\aggregate.py file where the latest_duration key is being updated incorrectly which is causing training for the newly aggregated datasets to fail.

This occurs when aggregating 3+ datasets but was also occuring when video data was exceeding the "DEFAULT_VIDEO_FILE_SIZE_IN_MB" size.

This PR fixes this by removing the unused episode duration and properly updating the latest_duration key data by setting it equal to the timestamps_shift_s instead of adding the ever increasing timestamps_shift_s variable.

We have also set the latest_duration to 0 when rotating to a new file / chunk as the episode metadata parquet file requires the timestamp for a new file to begin at 0.

The existing tests pass here

pytest tests/datasets/test_aggregate.py

A robust way to confirm this fix works is to get 3 datasets and combine them and test to ensure the standard training script trains properly.

A second test would be to get datasets such that the video data pushes past the DEFAULT_VIDEO_FILE_SIZE_IN_MB variable

[michel-aractingi]

Hey @nicholas-maselli you're right there is a bug in aggregate and we have a fix similar to what you do in this PR #2100

[nicholas-maselli]

Hey @nicholas-maselli you're right there is a bug in aggregate and we have a fix similar to what you do in this PR #2100

Oh excellent! I actually just pushed an additional fix here that fixes another rotating episode bug.

Do you need help with any dataset tools? I would love to help if there are any timelines for the release. I have several extremely large datasets I can test all your tools on if you would like =)

[michel-aractingi]

That would be great @nicholas-maselli ! We're planning to release it tomorrow but still it would be great if you test it and report or push any fixes that you find.