Physical Intelligence's PI05 robot can clean a home it has never seen — 50% of the time

Summary

Physical Intelligence has released PI05, a model that can send a robot into a home it has never seen and complete long-horizon tasks — cleaning a kitchen, tidying a bedroom — without any environment-specific training data. The success rate sits at roughly 50% to 80% depending on the task, up from what the company's co-founder describes as effectively zero under the previous state of the art, which required collecting data in the exact environment where the robot would be demonstrated.

The breakthrough is generalization. Before PI05, showing a credible robotics demo meant training on the specific room where the demo would happen. Now the robot arrives somewhere new, builds a working understanding of the layout and objects, and attempts the task from scratch. It doesn't always succeed, but it no longer needs a rehearsal.

Why simulation hasn't solved this

Simulation has worked well for locomotion — teaching a robot how to walk across varied terrain — because the hard problem there is modeling the robot's own body, which you do once and then randomize across environments. Manipulation is different. Every object the robot touches has different physical properties, and simulating those interactions accurately enough to transfer to the real world has proven nearly intractable. The data that matters most for dexterous manipulation — how a sponge compresses, how a sheet deforms — is also the data least likely to exist anywhere on the internet or be easy to describe in language.

Physical Intelligence's answer is to go straight to real-world collection. The team buys objects, brings them into varied physical spaces, and records interactions directly. One finding that surprised the team: the bulk of the training data for PI05 is not footage of mobile manipulators in homes. That category represents a small percentage of the total. Data from static arms in office settings and video from the internet both contributed, and the model learned to integrate them. That cross-source generalization suggests the scaling path may not require physically visiting tens of thousands of homes — a question the team describes as existential for the approach before PI05 resolved it.

End-to-end learning is already the architecture

Everything Physical Intelligence has shipped is fully end-to-end: raw camera input and a handful of other sensors in, motor actions out. There is no hand-coded control layer sitting in between. The company's view is that this isn't a design choice so much as a constraint — the world is too complex and variable to be described in explicit rules, the same way rule-based chatbots failed before large-scale language model training took over.

What PI05 does not yet have is a confirmed scaling law in the LLM sense — a clean relationship where more compute predictably produces better robot behavior. The team is actively searching for that recipe. In the meantime, they note that even a modest fleet of robots generates data at volumes comparable to what large language model labs consume for training, with no ceiling analogous to the finite size of the indexed web.

Consumer deployment is years away

The company is structured as a research lab, not a consumer product company. The threshold for home deployment, in the team's framing, is closer to 98% to 99% reliability, not 50%. When failures do happen at that stage, the model is a human teleoperator stepping in to finish the task — similar to how supervised autonomy worked in early robotics deployments. The overnight use case the team finds most compelling: laundry folded, meals prepped, house tidied while occupants sleep, so failures are low-stakes and don't interrupt daily life.

Tariffs and supply chain

On tariffs, the team's view is that the industry is insulated for now because almost no robotics company is in scaled commercial production — spending is concentrated in R&D rather than volume manufacturing. The practical effect is that there is time to develop US-based actuator suppliers and other critical supply chain components before the cost exposure becomes material.

The self-driving analogy

Physical Intelligence is skeptical of the robotics-equals-self-driving framing, not because it's wrong but because it leads investors down incorrect assumptions. The team points to Waymo outperforming Tesla despite having far fewer vehicles on the road, as a reminder that passive data scale doesn't automatically translate to capability. More pointedly, they remind prospective investors that most robotics demos circulating online are either teleoperated or show locomotion feats like backflips — a fundamentally easier problem than dexterous manipulation. Their honest timeline: a 15-year arc, similar to self-driving, is a realistic prior. Their stated greatest competition is science itself, not any specific rival company.