Two Released Checkpoints Appear Identical

[ChushanZhang]

Hi @LuFan31 @YangS03,

We noticed that the two released RoboTwin post-training checkpoints appear to be identical files.

Evidence

MD5 hash comparison:

4d7d1cbf55b45684f35a66d80c880643  lingbot-vla-4b-posttrain-robotwin/model.safetensors
4d7d1cbf55b45684f35a66d80c880643  lingbot-vla-4b-depth-posttrain-robotwin/model.safetensors

Weight-by-weight comparison:

Total keys: 1555
Keys with any difference: 0
Max diff: 0.0
Verdict: IDENTICAL checkpoints

Config files: also identical (config.json and lingbotvla_cli.yaml diff returns nothing).

No depth-specific weights found:

depth_keys = [k for k in keys if 'depth' in k.lower() or 'align' in k.lower()]
# Result: [] (empty)

The lingbot-vla-4b-depth-posttrain-robotwin checkpoint contains 1555 keys, all of which are VLM + Expert weights — no depth_align_head or depth-related parameters.

Questions

1. Is this intentional? (e.g., depth knowledge was fully distilled into the VLM weights, so no separate depth_align_head is needed at inference time?)
2. Or was the wrong file uploaded to HuggingFace for the depth variant?
3. If the depth checkpoint is different from the non-depth one, could you re-upload the correct weights?

We are trying to reproduce the w/ depth results from Table S7 (e.g., open_microwave: 92%, click_bell: 97%). Since both checkpoints are identical, we are currently getting the same results for both variants.

Thank you!

Originally posted by @ChushanZhang in #24

[ChushanZhang]

The same SHA256 demonstrate this as well:

[LuFan31]

Hi @ChushanZhang,

Thank you for your detailed investigation and for pointing this out. We sincerely apologize for this oversight.

We have now updated the lingbot-vla-4b-depth-posttrain-robotwin weights on Hugging Face with the correct checkpoint. Please feel free to download and test it again.

This depth version was also reproduced using the same setting in the report, we recommend comparing your results against the reproduction data we provided in #24, considering the performance fluctuations on individual tasks may result from the mix-training on 50 tasks, just as revealed in #24.