delta&absolute training; stats computing

[floating-petal]

Thank you for the great work on open source.

1. May I ask if there are plans to support delta and absolute training?
2. Currently, the stats are directly used from the pre-calculated /meta/episodes_stats.jsonl in the dataset. Will there be support for calculating stats in advance? For example, if I change the data from delta to absolute, I will need to recalculate the stats.

[shuheng-liu]

Hi @floating-petal , thanks for your interest in OpenTau. Currently, we use whatever dataset the user brings.

However, we can provide a script to convert your dataset. Thanks to the overridable standard data format, you can include an extra column in your datasets and decide which one to use during training. We will keep you updated once we finish the script and write up a tutorial.

(From the users' perspective, I don't think it's efficient to convert between absolute and delta actions on-the-fly. Some datasets can be very large and precomputing the min/max/avg requires iterating through the dataset once, and computing std requires a 2nd iteration.)

[floating-petal]

@shuheng-liu Thank you for your response, your suggestion is reasonable.
Additionally, do you have plans to publish some post-training metrics for using your code on open-source benchmarks (based on PI05_base weights)? For example, Robotwim2.0.

[WilliamYue37]

@floating-petal We do not have any plans to post simulation benchmark results beyond LIBERO results. We are planning to release real world experiments results, videos, and models in the future.