Request for π0.5 checkpoint trained on DOMINO dataset

[wangerforcs]

Hi,

Thanks for your excellent work on DOMINO and PUMA. I have successfully set up the environment and been able to run the provided PUMA checkpoint and reproduce the inference pipeline on the DOMINO benchmark.

According to the paper, you evaluate several baselines including π0.5 on the DOMINO dataset, where it achieves around 9.6% success rate in dynamic settings.

I would like to further reproduce and compare with these baselines, and was wondering:

Is the π0.5 checkpoint trained (or fine-tuned) on the DOMINO dataset publicly available?
If so, could you provide a download link or instructions to obtain it? If not, do you have plans to release these baseline checkpoints?

Having access to the π0.5 checkpoint would greatly help in benchmarking and validating results against your reported numbers.

Thanks again for your great work and for releasing this project!

Best regards

[FangHeng]

Hi, Thank you for your interest in DOMINO and PUMA, and for reproducing the inference pipeline on the DOMINO benchmark.

We are happy to share the pi0.5 checkpoint used in our benchmark. It is available here: Google Drive link. We hope this will help with reproducing the results reported in the paper and facilitate comparison with the baselines.

Please let us know if you have any further questions.

[wangerforcs]

Hi, @FangHeng

Thank you for sharing the checkpoint.

I tried to load it, but I ran into a couple of issues:

1. The checkpoint package does not contain the required assets directory.
2. The file pi05_domino/20260216-pi05-domino_dynamic_task-dynamic-35task/30000/params/ocdbt.process_0/d/0d560b50a3289841e14033520cf4432d appears to be 0 bytes.

Because of this, I am currently unable to load the checkpoint successfully.

ValueError: OUT_OF_RANGE: Error reading "params.PaliGemma.llm.embedder.input_embedding.value/1.0" in OCDBT database at local file "pi05_domino/20260216-pi05-domino_dynamic_task-dynamic-35task/30000/params/": Requested byte range [0, 414751475) is not valid for value of size 0 [source locations='tensorstore/kvstore/batch_util.h:340\ntensorstore/internal/cache/kvs_backed_cache.h:208']

Could you please clarify whether this is expected? If so, what is the correct way to use this checkpoint? In particular, could you share:

• the exact configuration needed to load it,
• any missing files or directories that should accompany the checkpoint

If possible, it would be very helpful if you could also provide a minimal example command or config for loading this checkpoint correctly.

Thanks again for your help.

[FangHeng]

Thanks for reporting this, and sorry for the confusion.

It turns out that some files were accidentally missing when I uploaded the checkpoint package, so the version at the link was incomplete. I have now updated the files at the same GoogleDrive link.

Could you please re-download the checkpoint from the same link and try again? The updated package now includes the missing checkpoint data, including the assets directory required for loading.

Sorry again for the inconvenience, and thank you for catching this.

[wangerforcs]

Hi @FangHeng,

Thanks for the quick fix and for updating the checkpoint!

I re-downloaded the package from the link, and the checkpoint now loads successfully on my side.

Best regards

[wangerforcs]

Hi @FangHeng,

Thank you again for sharing and updating the dynamic-task π0.5 checkpoint. I would also like to ask whether you could share a π0.5 checkpoint trained on the static environment as well.

I am currently trying to compare results under both dynamic and static settings, so having access to the static-data-trained checkpoint would be extremely helpful.

If such a checkpoint is already available, could you please provide the download link?

Thanks again for your great support!

[yqzhou886]

Hi @FangHeng,

Thank you again for sharing and updating the dynamic-task π0.5 checkpoint. I would also like to ask whether you could share a π0.5 checkpoint trained on the static environment as well.

I am currently trying to compare results under both dynamic and static settings, so having access to the static-data-trained checkpoint would be extremely helpful.

If such a checkpoint is already available, could you please provide the download link?

Thanks again for your great support!

Hi, I also tried the dynamic checkpoint mentioned here, but it completely fails on my side. None of my runs succeeded, and the success rate is 0%. For example, AdjustBottle always fails.

Did you observe the same behavior, or were you able to get any successful runs? I’m trying to confirm whether this is expected or likely a configuration issue on my side.

[wangerforcs]

Hi @FangHeng,
Thank you again for sharing and updating the dynamic-task π0.5 checkpoint. I would also like to ask whether you could share a π0.5 checkpoint trained on the static environment as well.
I am currently trying to compare results under both dynamic and static settings, so having access to the static-data-trained checkpoint would be extremely helpful.
If such a checkpoint is already available, could you please provide the download link?
Thanks again for your great support!

Hi, I also tried the dynamic checkpoint mentioned here, but it completely fails on my side. None of my runs succeeded, and the success rate is 0%. For example, AdjustBottle always fails.

Did you observe the same behavior, or were you able to get any successful runs? I’m trying to confirm whether this is expected or likely a configuration issue on my side.

I tested AdjustBottle as well. On my side, the success rate is around 50%, which is fairly close to what was reported in the paper.

[yqzhou886]

Hi @FangHeng,
Thank you again for sharing and updating the dynamic-task π0.5 checkpoint. I would also like to ask whether you could share a π0.5 checkpoint trained on the static environment as well.
I am currently trying to compare results under both dynamic and static settings, so having access to the static-data-trained checkpoint would be extremely helpful.
If such a checkpoint is already available, could you please provide the download link?
Thanks again for your great support!

Hi, I also tried the dynamic checkpoint mentioned here, but it completely fails on my side. None of my runs succeeded, and the success rate is 0%. For example, AdjustBottle always fails.
Did you observe the same behavior, or were you able to get any successful runs? I’m trying to confirm whether this is expected or likely a configuration issue on my side.

I tested AdjustBottle as well. On my side, the success rate is around 50%, which is fairly close to what was reported in the paper.

Thanks for the confirmation and quick reply! That’s helpful to know. I’ll re-check my configuration

[FangHeng]

Hi @FangHeng,

Thank you again for sharing and updating the dynamic-task π0.5 checkpoint. I would also like to ask whether you could share a π0.5 checkpoint trained on the static environment as well.

I am currently trying to compare results under both dynamic and static settings, so having access to the static-data-trained checkpoint would be extremely helpful.

If such a checkpoint is already available, could you please provide the download link?

Thanks again for your great support!

Hi @wangerforcs, sorry, I missed your comment on this closed issue. Do you still need the π0.5 checkpoint trained on the static setting? If so, I will organize it and update the link here in the next few days.

For reference, our π0.5 models were trained on the corresponding 35 RoboTwin2.0 tasks, following the official RoboTwin training/evaluation instructions. In practice, how to best use the available data, whether static or dynamic, is still a promising open question. We have also observed some interesting findings across different data compositions, so we would encourage everyone to experiment with different ways of organizing and mixing the data. This may help elicit more robust, generalizable, and dynamics-aware policies.

[yqzhou886]

Hi @FangHeng,
Thank you again for sharing and updating the dynamic-task π0.5 checkpoint. I would also like to ask whether you could share a π0.5 checkpoint trained on the static environment as well.
I am currently trying to compare results under both dynamic and static settings, so having access to the static-data-trained checkpoint would be extremely helpful.
If such a checkpoint is already available, could you please provide the download link?
Thanks again for your great support!

Hi @wangerforcs, sorry, I missed your comment on this closed issue. Do you still need the π0.5 checkpoint trained on the static setting? If so, I will organize it and update the link here in the next few days.

For reference, our π0.5 models were trained on the corresponding 35 RoboTwin2.0 tasks, following the official RoboTwin training/evaluation instructions. In practice, how to best use the available data, whether static or dynamic, is still a promising open question. We have also observed some interesting findings across different data compositions, so we would encourage everyone to experiment with different ways of organizing and mixing the data. This may help elicit more robust, generalizable, and dynamics-aware policies.

May I ask if there is any recommended guide or reference setup for evaluating the π0.5 checkpoint? After downloading the released weights, I observed obvious jittering of the robot arm during evaluation, while the PUMA policy runs smoothly under the same environment. Could this be related to π0.5-specific settings, such as action normalization, action chunk size, inference horizon, or other evaluation configurations?

[FangHeng]

Hi @FangHeng,
Thank you again for sharing and updating the dynamic-task π0.5 checkpoint. I would also like to ask whether you could share a π0.5 checkpoint trained on the static environment as well.
I am currently trying to compare results under both dynamic and static settings, so having access to the static-data-trained checkpoint would be extremely helpful.
If such a checkpoint is already available, could you please provide the download link?
Thanks again for your great support!

Hi @wangerforcs, sorry, I missed your comment on this closed issue. Do you still need the π0.5 checkpoint trained on the static setting? If so, I will organize it and update the link here in the next few days.
For reference, our π0.5 models were trained on the corresponding 35 RoboTwin2.0 tasks, following the official RoboTwin training/evaluation instructions. In practice, how to best use the available data, whether static or dynamic, is still a promising open question. We have also observed some interesting findings across different data compositions, so we would encourage everyone to experiment with different ways of organizing and mixing the data. This may help elicit more robust, generalizable, and dynamics-aware policies.

May I ask if there is any recommended guide or reference setup for evaluating the π0.5 checkpoint? After downloading the released weights, I observed obvious jittering of the robot arm during evaluation, while the PUMA policy runs smoothly under the same environment. Could this be related to π0.5-specific settings, such as action normalization, action chunk size, inference horizon, or other evaluation configurations?

Hi @Roswellii, thanks for the question.

We did not make any π0.5-specific adjustments — both fine-tuning and evaluation strictly follow the official RoboTwin 2.0 openpi instructions with their default settings (action normalization, chunk size, inference horizon, etc.): https://robotwin-platform.github.io/doc/usage/Pi05.html. so we'd suggest first checking that your eval pipeline matches the official defaults.

As for the jittering: this is largely expected rather than a misconfiguration. Dynamic manipulation is intrinsically harder — a single-frame policy replans against a moving target without temporal context, so consecutive action chunks can disagree, producing less smooth trajectories than a history-aware architecture like PUMA. We observed similar qualitative behavior in our own evaluations. That said, jitter alone should not cause a 0% success rate, so a config check against the official instructions is still worthwhile.

More broadly, how to best adapt single-frame VLAs to dynamic settings — static/dynamic data composition, training strategies, chunk length vs. replanning frequency — remains an open question. We encourage further exploration on DOMINO.

Hope this helps!

[tonyzhu163]

@Roswellii @FangHeng

Hi all,

I've been having similar issues replicating the π0.5 results with the shared checkpoint (0/50 on AdjustBottle dynamic vs the paper's 52%, and 2/50 static vs 70%). Was wondering if theres any setup info that I am potentially missing. I've verified that I'm using the same checkpoint from gdrive and that the code is running on the exact uv.lock environment, and I've also aligned against robotwin settings. Please let me know if providing any more information would be helpful. Really appreciate it.

Thanks,

[FangHeng]

Hi @tonyzhu163,

Thanks for reporting this. As @wangerforcs confirmed earlier, they achieved ~50% on AdjustBottle (dynamic) with the same checkpoint, so the model itself should be functional — a 0% success rate likely indicates a configuration mismatch.

To help diagnose, could you share the exact eval command you are running (especially train_config_name, task_config, and seed), and describe what you observe during evaluation? For example, is the robot moving at all, jittering in place, or moving in wrong directions? This helps distinguish between different types of failures.

The most common cause we've seen is a mismatch in adapt_to_pi. The pi05_aloha_full_base config uses adapt_to_pi=True by default, while the pi0 RoboTwin configs explicitly set adapt_to_pi=False. If you are using a pi0-style config or have manually set adapt_to_pi=False, the joint angles and gripper positions will be completely wrong, leading to 0% success. Please make sure you are using pi05_aloha_full_base as the train_config_name without modifying this setting. Also verify that the assets/ directory exists under your checkpoint path and contains the norm_stats files.

As a quick sanity check, try printing the first few predicted actions to see if they look reasonable (not NaN or extremely large). Please let us know what you find!

---