Questions about action space definition of DM0

[Deephome]

Thank you for your great work of DM0!

I am curious about the action space definition of DM0 （ EEF, joint space, or hybrid）, which is not clearly clarified in your technical report. In the source code of DM0, the action_dim=32. In DM0's RoboChallenge inference code, it seems to use joint space for ALOHA, and use EEF for single arm robots. would you like to provide more details on this point?

[dexmal2025]

Thanks for your pointing this question ! For DM0 pretraining, we use hybrid action representation, including both EEF and joint. While for RoboChallenge finetuning, we use joint for ALOHA and EEF for other single arms to produce better performance. In our daily experiments, we found joint space is more suitable for ALOHA arms. The action_dim is set to 32 for max-length of different embodiedments.