MolmoAct2 says open robots just outscored GPT-5
AI2 dropped MolmoAct2 on May 4. It's an open-weight action reasoning model for real robots, and the embodied-reasoning component MolmoER beats GPT-5 and Gemini Robotics ER 1.6 across 13 benchmarks. Not in a paper-only sense. They tested on 7 simulation and real-world setups, the most extensive empirical study of any open VLA so far.
The interesting trick is MolmoThink. Instead of replanning the whole scene every step, it adaptively reprocesses only the regions that changed. Fraction of the prior latency, geometric understanding intact. This is the part Gemini Robotics ER 1.6 doesn't do for free. Plus they shipped MolmoAct2-BimanualYAM, 720 hours of teleoperated trajectories, the largest open bimanual dataset to date.
So an open AI2 lab just out-benchmarked Google DeepMind's flagship robot reasoning model and gave the dataset away. Read alongside the Standard Intelligence raw-video bet last week and it's the same story at a different layer. The closed labs do not have a structural lead in embodied reasoning anymore. The open weights and the open data both showed up the same week.
Paper at https://arxiv.org/abs/2605.02881.
← Back to all articles
The interesting trick is MolmoThink. Instead of replanning the whole scene every step, it adaptively reprocesses only the regions that changed. Fraction of the prior latency, geometric understanding intact. This is the part Gemini Robotics ER 1.6 doesn't do for free. Plus they shipped MolmoAct2-BimanualYAM, 720 hours of teleoperated trajectories, the largest open bimanual dataset to date.
So an open AI2 lab just out-benchmarked Google DeepMind's flagship robot reasoning model and gave the dataset away. Read alongside the Standard Intelligence raw-video bet last week and it's the same story at a different layer. The closed labs do not have a structural lead in embodied reasoning anymore. The open weights and the open data both showed up the same week.
Paper at https://arxiv.org/abs/2605.02881.
Comments