MolmoPoint from Ai2 was a recent foundation model release that feels criminally underrated to me, while not a robotics model it has great potential for various use cases labels are great even for wrist view!

I do feel like when stacking the block on the tower, the model should point at the tower and not the block we're holding. Also not every task I tried had labels as clean as these. but this feels like a user error fixable via prompting, not an issue with the model

Grounding models are the missing piece for agents that need to interact with actual UIs. Most browser agents right now rely on DOM parsing or OCR, both super brittle. Something that just points to 'the submit button' from raw pixels would be massive for non-web automation. Curious if anyone's tried plugging this into an agent loop yet.

@allen_ai very cool! What would happen if scenes are more cluttered?

@allen_ai I haven't tested it that much yet but on some less obvious tasks and more cluttered scenes it didn't do as well. but that could also come down to inefficient prompting

@allen_ai OH YES. Wrist view targeting that actually works. This is unicorn-level innovation for robot vision.

@allen_ai Wrist view accuracy is what sets this apart. Grounding models finally getting practical!