Paper Club #5 (Robot Learning Collective)

​A lot of impactful robotics research today happens inside companies. As a result, we often see the polished success story, but not the full picture: what worked, what failed, and which decisions actually mattered.

That is why MolmoAct 2 is especially interesting.

It is a rare example of a large, open work on building a strong VLA policy from pretraining to fine-tuning. The authors release the data, code, and checkpoints, explain the training procedure, and include detailed ablations.

This makes the paper very valuable for anyone who wants to understand how modern robot policies are actually trained, not just what the final benchmark numbers look like.

On June 15, we will go through the work in detail and discuss the results, design choices, and what we can learn from them.

Join us if you are interested in robotics, VLA models, or practical details behind training robot policies.

More details in our Discord channel.