libero checkpoint (s)

[kevinlee1202]

Hi, thanks for the great work and for open-sourcing the code!

The ModelScope repo (hanyangyu1021/PoseVLA-robotwin) currently only contains
the RoboTwin post-trained checkpoint. Since the paper also reports strong
results on LIBERO (96.0% average, Table III), would you consider releasing
the LIBERO post-trained checkpoint(s) as well?

We are studying VLA failure modes on LIBERO-based benchmarks, so being able
to evaluate PoseVLA directly on LIBERO would be extremely helpful.

A few related questions:

1. Could you also share the LIBERO data-conversion and post-training configs
   (the repo currently documents RoboTwin only)? That would let us reproduce
   the LIBERO fine-tuning even if releasing the weights takes time.
2. If more convenient, the clean spatial pre-training checkpoint (before any
   downstream post-training) would serve the same purpose, since we could
   run the LIBERO post-training ourselves with the configs from (1).

Thanks a lot!

[hetolin]

Hi! Thanks for your interest.

We've released the Stage-1 pretrained checkpoint (before any downstream post-training). You can directly use it to post-train on LIBERO following a similar pipeline as docs/POSTTRAIN.md.

We haven't organized the LIBERO-specific configs yet — it's on our roadmap. In the meantime, you could also try the RoboTwin checkpoint for a quick evaluation.