[Question] Design trade-offs and differences vs. LeRobot

[author31]

Hi TRI Team,
I’ve been following the VLA Foundry repo and noticed you’ve implemented your own training stack and dataset structure (WebDataset) rather than using existing frameworks like LeRobot.
Out of curiosity and for my own learning, could you share the primary limitations you found in existing stacks that necessitated a custom implementation?

[arpitg1304]

I guess there are couple of reasons for using WebDataset tars

1. S3 streaming for multi-node : pipes tars straight from s3 - no local copy, lerobot would be slow when you are scaling to many nodes on sagemaker
2. The framework also trains LLM/VLM so a single format across modalities is chosen as it is not just robotics-specific
3. Shard-level shuffle seed is checkpointed for exact replay — easier to do at the shard layer than over LeRobot's episode index.

[richardrl]

From my own experience LeRobot has horrific IO numbers. Not sure if it's due to using NFS, but it's very very slow with large datasets with long episodes.