issues when running training policy

[ZoranGong]

hi,

thanks for your awesome work. I'm trying to reproduce forcevla but I find the current code is not complete. When running the training policy, the problem occurs: "Config 'forcevla_lora' not found". Also, the options for train.py are unrecognized.

Can you help me solve the above problems?

[nhathout]

I am also reproducing forcevla and managed to successfully. I can try to scrape my logs to see if I ran into that issue, which I am sure I did, and then I can let you know soon what my fix was whenever I find it.

[ZoranGong]

I am also reproducing forcevla and managed to successfully. I can try to scrape my logs to see if I ran into that issue, which I am sure I did, and then I can let you know soon what my fix was whenever I find it.

If you're willing to share your reproduced repo, that will be great for the community!! For now, we find there are several key parts missing in the official repo (including config files), and authors didn't response our emails yet... Your successful experience will be a great help!

[Raunak-Chakraborty]

True. I too am using this for my thesis. It would be really nice if you could share what worked for you. Another useful fork that I found was from tshiamor. Also check that out.

[ZoranGong]

we are blocked at the beginning, since the it lacks core files for training : (

[nhathout]

Hey all, glad this is getting traction. A few things:

1. The forcevla_lora config error comes from a broken config definition in config.py — the field wiring is incomplete in the official release. It's not a missing file, it's in there but doesn't work as-is.

2. The unrecognized train.py arguments stem from the same area — the multi-dataset repo_id handling wasn't fully implemented.

3. I've put together a fork with the patches that got me to a working baseline: https://github.com/nhathout/ForceVLA/tree/thesis-forcevla-edits.

Also, I took a look at tshiamor's fork mentioned above, and it definitely looks like it could be useful.

[Raunak-Chakraborty]

Hi guys, I am trying to use forceVLA for peg-in-hole insertion for my Master's Thesis. How many hours of data did you collect for fine-tuning forceVLA? I am doing data collection in real robot (doosan m1013) using spacemouse teleoperation. Since data collection is a slow & cumbersome affair, I wanted to know how many hours/episodes of data need to be collected. Also, are any of your datasets on hugging face?