MolmoAct

MolmoAct is a 7-billion-parameter vision-language-action model from the Allen Institute for AI, released in August 2025. Built on the Molmo VLM backbone, it adds an action-reasoning layer that draws explicit trajectory lines over what it sees, so its decisions can be inspected and even edited before the robot moves. The team released weights, training code, and the full action-reasoning dataset.

On standard robotics benchmarks the authors report it outperforming baselines including NVIDIA's GR00T N1, Physical Intelligence's pi-zero and pi-zero-FAST, RT-1, and TraceVLA, with a paper-cited 97.2 percent average success on LIBERO. It is fully self-hosted with no official API, which makes it a research and robotics-team tool rather than a consumer product.

Paper-reported 97.2% average success on LIBERO (figures carried by community repros, e.g. OpenRAL rskills, citing Ai2); outperforms GR00T N1, pi-zero/pi-zero-FAST, RT-1, and TraceVLA on LIBERO and SimplerEnv; higher Elo than baselines in arena-style human evaluations (all Ai2 claims, not independently reproduced here)