A robot brain learned to use an air fryer.

Unexpected skill from a general model

Physical Intelligence, a two-year-old San Francisco robotics company, published research on 16 April 2026 showing a new model that can direct robots to perform tasks it was never explicitly trained on. The model, called π0.7, surprised the research team by synthesising fragments of prior experience and web-scale pretraining into workable behaviour.

So far, teams have had to shoot footage for each task, train a bespoke model, and then repeat that cycle every time a new job appears. π0.7 tries something different: it learns broad patterns from huge datasets, then pieces those skills together to tackle problems it hasn't seen before. That approach points toward what Physical Intelligence calls a general-purpose robot brain: point it at an unfamiliar job, give plain-language coaching, and watch it work.

How the air fryer test unfolded

The paper highlights a striking demonstration. The team presented the robot with an air fryer — an appliance the model had barely encountered during training. In the company’s dataset, there were only two relevant clips: one of a different robot nudging an air fryer closed, and another from an open dataset showing an unrelated robot placing a plastic bottle inside an air fryer following human instruction.

Despite that sparse exposure, π0.7 made a reasonable attempt at using the appliance to cook a sweet potato with no coaching. When a human gave step-by-step verbal instructions — much like explaining a task to a new employee — the robot completed the task successfully.

The research team says the result caught them off guard.

"Once it crosses that threshold where it goes from only doing exactly the stuff that you collect the data for to actually remixing things in new ways," said Sergey Levine, co-founder of Physical Intelligence and a UC Berkeley professor, "the capabilities are going up more than linearly with the amount of data."

Compositional generalisation, explained

At the core of the claim is a capability called compositional generalisation. In plain terms, that means the model can combine distinct skills it learned in different settings and apply the mix to an unfamiliar situation. A robot that knows how to open containers and how to move objects might be able to cook with an appliance it has never been taught to use by recombining those basics.

Physical Intelligence says π0.7 leverages both robotic demonstrations and large-scale web pretraining. Web data teach the model general visual and language patterns; real robot demonstrations show how to turn those patterns into actual movements. The result, the company argues, is a model that doesn't simply recall a memorised script for each task but that builds flexible plans from components.

"It’s very hard to track down where the knowledge is coming from, or where it will succeed or fail," said Lucy Shi, a researcher at Physical Intelligence and a Stanford computer science PhD student. She added that isolating the provenance of a particular behaviour remains difficult.

Mastheads

Why researchers are paying attention

The robotics field has long chased a generalist controller that can adapt on the fly.

Specialist models work well, but they don't scale: every new tool or workflow calls for fresh data and retraining. If other labs can reproduce these results, this could be a turning point—similar to when big language models suddenly showed unexpected new abilities as they scaled up.

Those emergent abilities can compound. Levine’s comment about capabilities rising faster than data suggests companies may get more flexible robots without a proportional rise in labelled demonstrations. In practice, that could mean robots entering new environments and improving through simple verbal coaching rather than months of data collection.

That shift would alter how teams deploy robots in warehouses, factories, labs and homes. Instead of long retraining cycles, operators could point and speak, then watch a robot try, correct and repeat.

Technical and practical hurdles

The research paper also lays bare new challenges. If a model's knowledge is distributed across web pretraining and scattered robotic examples, it becomes hard to predict failure modes. A robot might succeed in one kitchen and fail in another for reasons that aren't obvious from the training data. Troubleshooting requires careful probes and new benchmarks.

There are safety worries: if the model stitches behaviours together the wrong way, it might try dangerous actions near people or hot equipment. Engineers will need guardrails that constrain exploration when a robot is near fragile objects, heat sources or people. Those constraints must balance flexibility with predictable behaviour.

And crucially, outside teams will need to reproduce the experiments before we can trust the claims. The company’s paper includes demonstrations and dataset descriptions, but widespread adoption depends on independent replication, standardised tests and community scrutiny.

Researchers will want to know how π0.7 performs across a range of tools, materials and lighting conditions, and how brittle its solutions are when a single element changes.

What this means for robot development

If the results scale, developers could shift investment from collecting task-by-task datasets to building larger, more diverse foundations and designing better in-the-loop coaching methods. Human operators would act less like data gatherers and more like supervisors who teach and refine skills on site.

That model of deployment reduces friction for applications where customisation is the norm. Hospitals, catering kitchens and small manufacturers often need robots tuned to narrow workflows. A generalist brain that can be guided by staff in plain language could lower the cost and time to adoption.

At the same time, companies and labs will have to develop new tooling for monitoring what the model is doing and why. Debugging an emergent behaviour is harder than fixing a buggy specialist. The field will likely see growth in interpretability tools, richer logging, and simulators designed to stress-test compositional failures.

Next steps and scrutiny

Physical Intelligence's paper presents promising demonstrations, but the team itself admits there are mysteries. The apparent ability to recombine sparse data fragments raises as many questions as it answers. Researchers will examine the datasets, attempt replication, and design benchmarks that stress compositionality across varied tasks.

For adopters, the immediate takeaway is cautious optimism. The technology points toward robots that can adapt faster and learn from simple human coaching. It also means engineers and safety teams will need to update testing practices to catch odd failures born from cross-domain synthesis.

Related Articles

"Once it crosses that threshold where it goes from only doing exactly the stuff that you collect the data for to actually remixing things in new ways," said Sergey Levine, "the capabilities are going up more than linearly with the amount of data."

This article was created with AI assistance.