MolmoAct 2 Code and Training Data Released, Integrated Into Hugging Face LeRobot

MolmoAct 2 code and training data have been released. The release includes fine-tuning scripts, all datasets used for training, evaluation rollouts and instructions for training the open-source MolmoAct 2 tokenizer. According to the article, MolmoAct 2 artifacts were downloaded more than 400,000 times in the weeks after their release. Inference and training have also been integrated into Hugging Face’s LeRobot.

According to the article, researchers can use the released training data and evaluation rollouts to assess and fine-tune the model, and adapt it to different hardware and tasks. The integration of inference and training into Hugging Face’s LeRobot is intended to reduce the need for teams using the platform to reconfigure their existing environments.

A key feature of this release is that it provides not only the weights needed to use the model, but also the training data, fine-tuning scripts, evaluation rollouts and instructions for training the tokenizer as a package. With data and evaluation processes available, researchers can examine the model’s outputs and access resources for adapting it to different hardware and tasks. However, releasing these materials does not guarantee that the same performance can be reproduced. The composition and terms of use of the data, as well as specific figures for evaluation results, cannot be confirmed from the information provided. The article points to a broader issue: in robot foundation models, releases that include data and sufficient training information are less common than those that provide weights alone. In that context, this release expands what is available with the model and shares process information that third parties can use to verify and modify it. However, the fact of release alone does not establish how reproducible the materials are in users’ environments or how much adjustment is needed to adapt the model to different robots and tasks. The article says that integrating inference and training into LeRobot is intended to let teams already using the platform incorporate the model into their existing environments. In addition to releasing training materials, the integration provides a connection to the environment where the model is used. However, the description does not quantify how much the integration changes setup procedures or operational workload. The number of artifact downloads indicates interest and the scale of access, but does not necessarily represent real-world deployments or performance improvements. Questions to examine going forward include the datasets’ contents and licenses, the tasks and conditions covered by the evaluation rollouts, third-party reproduction results, and adaptation outcomes on different hardware. Those details would help assess how the breadth of the release translates into practical verifiability and applicability to robots.