Physical Intelligence: building a foundation model for any robot

Executive overview

Most robotics companies are locked into a single hardware platform, requiring years of effort before deploying a single useful task. Physical Intelligence (Pi) trains one cross-embodiment model across many different robot platforms, so intelligence transfers rather than needing to be rebuilt per robot.

The result: tasks that required hundreds of hours of data collection last year can now be performed zero-shot. Real deployments are already running in laundromats and e-commerce warehouses — two years into the company's life, not five.

The GPT-1 moment for robotics is not a future event — it is beginning now, and the cost of building on top of it is falling fast.

Why robotics was hard — and what changed

• The three pillars of robotics are semantics, planning, and real-time control — all historically solved separately with brittle, hardware-specific code.
• Language models (SayCan, RT-2) first cracked semantics: common-sense knowledge transferred into robot planning without robot-specific data.
• RT-2 showed that a vision-language model fine-tuned on robot data could reason about entirely unseen objects and spatial relationships.
• Open X-Embodiment was the first large-scale training run across 10+ robot platforms — and the generalist model outperformed per-platform specialists by 50%.
• Even a single robot platform drifts over time through hardware and software changes, making old data less reusable; a diverse fleet is actually easier to generalise from.

The cross-embodiment scaling insight

• Training on heterogeneous robot data teaches the model an abstract concept of "how to control a robot," not "how to control this robot."
• Data scarcity in robotics is two problems: data generation (operationally expensive) and data capture (incentive and infrastructure gaps).
• Pi's approach: be ready to absorb data from many different robot types already in the field rather than manufacturing and scaling one proprietary platform.
• Emergent transfer properties are now visible — tasks zero-shot today required hundreds of hours of collection a year ago.

Cloud inference and real-time control

• Almost all of Pi's robot evaluations — including laundry folding and mobile manipulation — run with the model hosted in a remote data centre, not on-device.
• The robot queries a cloud API endpoint, sending images and language commands and receiving back action chunks in a high-frequency control loop.
• Real-time chunking: the robot pre-fetches the next action chunk while still executing the current one, hiding inference latency.
• This eliminates the need for expensive, fast-obsoleting onboard compute, and decouples hardware choices from model capability.

Current state of deployment

• Pi partnered with Weave (home laundry robots) to fold diverse, unseen clothing items in a real laundromat — built to working demo in roughly two weeks.
• Partnered with Ultra (logistics automation) to pack varied items into soft e-commerce pouches over a full eight-hour day in a live warehouse with minimal human intervention.
• A mixed-autonomy model — human corrects errors, robot improves — is sufficient today to reach economic break-even and justify scaling.
• Pi intentionally stays ignorant of each partner's hardware internals; the model parachutes into their data pipeline, which tests true generalisation.

The playbook for vertical robotics startups

• Identify a workflow with a clear insertion point — where does a robot make the biggest difference without redesigning the whole process?
• Use cheaper hardware; the model compensates for imprecise motion.
• Collect data, run real-deployment evaluations, build a mixed-autonomy system.
• Reach economic break-even before scaling unit count — historical robotics companies failed by scaling a money-losing deployment.
• The barrier to entry has dropped: scrappiness and customer understanding now matter more than 20 years of robotics expertise.

Infrastructure gaps and internal tooling

• When Pi started, no off-the-shelf tooling existed for large-scale robot data collection, annotation, evaluation management, or operational monitoring — the team built it all.
• This is a significant opportunity: teleoperation services, data collection pipelines, annotation services, and evaluation infrastructure are needed by every vertical robotics company.
• Pi runs a Claude-based pre-training monitor that babysits large training runs autonomously, remedying errors in real time — producing ~50% improvement in compute utilisation.
• The next frontier: an automated robotic research scientist that ingests multimodal failure data and proposes and tests hypotheses across the full training stack.