MolmoAct 2: An open foundation for robots that work in the real world
Source
allenai.org
Date
Context
MolmoAct 2 makes robot action generation substantially faster in Ai2's reported test. It connects a vision-and-LLMA large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.Full definition → specialized for spatial reasoning to a component that generates robot actions. In the LIBERO environment on one NVIDIA H100, the base model took about 180 milliseconds per action call, compared with 6,700 milliseconds for its predecessor. The release specifies that comparison.
Faster action generation still has an important control limit: the robot executes batches of 10–30 moves before running inferenceRunning a trained model to get answers — the phase where AI is actually used, as opposed to trained.Full definition → again, so it cannot adjust midway through a batch. The result shortens the model's planning pauses, while leaving continuous reaction to unexpected changes as a separate engineering challenge.
Terms in this piece · Glossary
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
