Missing files and training code needed to reproduce UniAct from this repo

[vkgo]

Hi UniAct authors,

Thanks for releasing the code! I’m trying to reproduce the UniAct text2action pipeline (text → motion tokens → decoder → tracking policy → MuJoCo) using this repository, and potentially retrain the T2A model with our own data. From checking the code, I think several key pieces are missing/unclear:

For inference (text2action)

1. The motion generator checkpoint directory checkpoints/motion-genorator/ only contains configs and model.safetensors.index.json, but is missing the actual weight shards:
   • model-00001-of-00002.safetensors
   • model-00002-of-00002.safetensors
2. server.py imports from fsq import FSQ, but there is no fsq module in the repo. Could you point to the correct implementation/package?
3. The checkpoint folder seems to miss tokenizer assets (e.g., vocab.json), so AutoTokenizer.from_pretrained() may fail in offline setups. What exact files should be included for the motion generator checkpoint?

For training (retraining T2A / motion generator)

4. Lack of training code for:

   • the tracking policy (RL)
   • the FSQ tokenizer/encoder (motion → token labels)

5. If we want to retrain only the motion generator while keeping decoder.pt + policy.pt unchanged, we need the exact FSQ token labels (or the tokenizer/encoder to generate them) to keep the token space identical.

Could you please provide links or instructions to obtain the missing ckpts/files and (if available) the training code?

Thanks again!

[physercoe]

the author seems to update the model files, and i use the fsq package from https://github.com/duchenzhuang/FSQ-pytorch and also change some codes, now the code(server and robot_client) could be run but not as expected: just fall down not walk.

[jnnan]

We're working on updating the repository with these missing components and expect to have everything in place within the next few days. Thanks you for your patience.

[vkgo]

Thx for the authors' excellent work and for codes. After the update, I have successfully run the language-to-action pipeline, but I still have two questions.

1. Is there a plan to make the splits for UA-Net and the ckpts for TMR publicly available?
2. It seems that the implementation code for the music/trajectory-to-action pipeline is still missing.