We are trying to adapt MiniCPM-RobotManip to a new robot arm for a specific manipulation task. The inference side is clear — vla_infer.py exposes the embodiment_id interface and action_head.max_num_embodiments for multi-embodiment support. However, the current repository only includes inference code; no training or fine-tuning scripts are provided for RobotManip.
Questions
-
Fine-tuning scripts: Is there a plan to open-source the training/fine-tuning code for MiniCPM-RobotManip? Even a minimal script covering data loading, optimizer setup, and checkpoint saving would be very helpful.
-
New embodiment registration: What is the recommended workflow for adding a new embodiment_id beyond the ones already in the released weights? Does this require retraining from scratch, or is there a fine-tuning path that only updates the action head for the new embodiment?
-
Data format: Could you document or provide a sample of the training data format expected by RobotManip (state shape, action shape, image resolution, etc.)? The inference script implies STATE_DIM = 80 and action shape (30, 80), but the mapping from real robot joints to this 80-dim space is unclear.
-
Data quantity guidance: Are there any ablation results or rule-of-thumb recommendations for how many demonstration episodes are needed to fine-tune on a new embodiment/task?
Any pointers — even a rough timeline or a link to an internal script — would be greatly appreciated. Thank you for the great work on this project!
We are trying to adapt MiniCPM-RobotManip to a new robot arm for a specific manipulation task. The inference side is clear —
vla_infer.pyexposes theembodiment_idinterface andaction_head.max_num_embodimentsfor multi-embodiment support. However, the current repository only includes inference code; no training or fine-tuning scripts are provided for RobotManip.Questions
Fine-tuning scripts: Is there a plan to open-source the training/fine-tuning code for MiniCPM-RobotManip? Even a minimal script covering data loading, optimizer setup, and checkpoint saving would be very helpful.
New embodiment registration: What is the recommended workflow for adding a new
embodiment_idbeyond the ones already in the released weights? Does this require retraining from scratch, or is there a fine-tuning path that only updates the action head for the new embodiment?Data format: Could you document or provide a sample of the training data format expected by RobotManip (state shape, action shape, image resolution, etc.)? The inference script implies
STATE_DIM = 80and action shape(30, 80), but the mapping from real robot joints to this 80-dim space is unclear.Data quantity guidance: Are there any ablation results or rule-of-thumb recommendations for how many demonstration episodes are needed to fine-tune on a new embodiment/task?
Any pointers — even a rough timeline or a link to an internal script — would be greatly appreciated. Thank you for the great work on this project!