AI2 open-sources MolmoMotion: 4B VLM forecasting 3D point trajectories with code and data
rsasaki0109 · x · 2026-10-03
AI2 has open-sourced the MolmoMotion repo on GitHub. MolmoMotion is a 4B vision-language model predicting 3D point trajectories from natural-language action instructions: given a short RGB observation history, user-specified 2D query points with initial 3D positions, and a language description, it forecasts each point's 3D trajectory up to 2 seconds ahead. The repo ships the autoregressive (AR) variant from the paper, the MolmoMotion-1M training corpus, and the PointMotionBench evaluation suite, plus data-generation and robotics modules. The motion prior transfers to robotics planning and motion-guided video generation.
More from Embodied
- Robotics startup Silica debuts affordable actuators with 10x force sensitivity — Scobleizer · 2026-10-03
- NUS, Tsinghua, PKU and 4 more universities open-source OpenWAM world-action foundation model — jiqizhixin · 2026-10-03
- Muse Gadget Ported to a Conference Badge Running MicroPython on RP2350 — alexandr_wang · 2026-10-03
- RNE: an open-source Rust robot-native engine for deterministic sim and policy evaluation — rsasaki0109 · 2026-10-03
- CMU lab spotted training a robotic arm to smack Pepper robot's butt — kamathsblog · 2026-10-03
- Self-sensing actuator senses forces down to 1g, 50x more precise than standard Chinese actuators — Scobleizer · 2026-10-03