Ego motion negative timestamps ?
There are some negative timestamps in the ego motion labels, for example for the clip 0c8731a8-9cd4-4f59-9603-165b1ed53e07, we have:
I'm loading with Pandas .read_parquet(), is this expected ?
Yes, this is intentional/expected -- we recognize that this is a bit unusual compared to the convention of an unsigned dtype (uint64 or otherwise) for timestamps, but we made the choice to provide all timestamps, e.g., camera frame timestamps, other sensor timestamps, egomotion, etc., for all clips relative to a start timestamp so that each 20s clip nominally comprises a period 0 -- 20_000_000 microseconds. Due to the asynchronous nature of various systems you'll find (slightly) negative timestamps quite often throughout the dataset. For egomotion specifically, we did intentionally always want to provide a bit of history (as you observe) running up to the main time segment, as well as a decent amount of future (typically ~2 minutes) in case it might be useful for mimicking routing intent.
In any case, thanks for the feedback! I'll collect this and any other comments/requests for clarification you & the broader community have during this initial release period to make an update pass through the docs.
Thanks ! Btw, great work and a big contribution for the community !