More capability from the models you already have.

Foundation models are transforming robotics. We build the metacognitive intelligence that knows how to use, combine and improve them.

Robotics AI has never had better models: VLAs, world models, vision and language models, skills and perception tools.

But a physical task rarely needs just one of them, and the real world never follows the script. Something has to pick the right capabilities, combine them, and adjust when things go wrong. That layer is missing from today’s robotics stack. Without it, capable models fail at work they should be able to do, and robots repeat their mistakes.

Motoniq builds it: metacognitive intelligence for the physical world. It knows what a task needs, which tool fits, and when something has gone wrong. It works across models and tools, making a robot’s existing capabilities more useful.

At its core is the Work Model: a metacognitive model trained for physical work. It learns from every task and gets better with each one.

Unlocking existing VLAs

Same model, no retraining.

Vision-language-action models carry broad, learned manipulation behavior, but individual tasks often demand something the policy does not have: a force-controlled contact, a precise fitting strategy, or the right context at the right moment.

We kept the existing policies (π0 on a physical robot, and LIBERO-finetuned π0 and π0.5 in simulation) and added Motoniq’s Work Model around them. Policy weights were left unchanged.

In our precision-fitting experiments, the baseline π0 policy succeeded in 0 of 10 trials. With orchestration and additional task information, the system succeeded in 9 of 10 trials.

Success rose on all six charted tasks, by 4 to 90 percentage points. The largest effects appeared where the base policy lacked a needed capability outright.

These results show how orchestration, task information and specialist capabilities can improve the performance of existing VLAs without changing their policy weights.

Prediction is not work.

Same model, made more useful.

World models are getting better at predicting the next state of a scene. But a plausible prediction is not a finished task, and contact-rich manipulation is where that gap shows.

We tested LeWM on its own, and LeWM with Motoniq’s Work Model on top, on PushT and OGB-Cube. Motoniq lifted LeWM from 90% to 96% on PushT and from 70% to 84% on OGB-Cube. Where LeWM alone drops the cube, the same model with Motoniq on top reaches the goal.

As world models improve, the Work Model gets better tools to use. Every advance in the field makes Motoniq stronger.

Same pattern as VLAs. We make world models more useful.