What happened
The Allen Institute for AI released MolmoMotion on June 17, an open model family for language-guided 3D motion forecasting. Given a video frame, a set of 3D query points and a written instruction, the model predicts where those points will move in 3D over the next few seconds. It is built on the Molmo 2 backbone and comes in autoregressive and flow-matching variants, with robotics planning as a primary use.
Ai2 openly released the weights along with the MolmoMotion-1M dataset, which contains 3D trajectories from 1.16 million videos across 736 motion types, and the PointMotionBench benchmark of 2,700 clips. Everything is available on Hugging Face. A policy built on MolmoMotion reached 76.3% pick-and-place success compared with a 56.0% baseline.
Why it matters
Open weights, dataset and benchmark together let others reproduce and extend the work rather than take results on faith. Language-guided 3D motion prediction is directly useful for robotics planning, where anticipating movement is essential.
The reported jump over baseline suggests the approach has practical value beyond the benchmark.
MintedBrain take
The fully open release is the standout here, since reproducible weights and data are worth more than a closed model with better numbers. Benchmark gains do not automatically transfer to messy real-world robots, so treat the pick-and-place figures as a starting point and validate on your own hardware.
Discussion
Sign in to comment. Your account must be at least 1 day old.