Physical Intelligence: the Bezos-backed company teaching robots to adapt
Image created by AI
Backed by Jeff Bezos and OpenAI, robotics startup Physical Intelligence is developing a universal AI "brain" capable of seamlessly transferring physical skills between entirely different robots.
The company’s foundation models are already deployed with partners to automate tasks like folding laundry and packing orders, while simultaneously providing a foundational software layer for third-party developers to power their own machines.
The viability of this cross-platform approach was recently highlighted in a breakthrough experiment. The company reported that its π0.7 model successfully folded clothing using two industrial UR5e arms after learning the task exclusively on much smaller, hobby-grade hardware. While the model had been trained on the industrial arms for other tasks, it had never seen a single folding demonstration on that setup, requiring the AI to autonomously adapt its movement strategy to complete the job. By proving that robotic intelligence can generalise across hardware, Physical Intelligence is paving the way for an ecosystem where businesses can easily adapt a single foundation model to any equipment.
The company is led by co-founder and chief executive Karol Hausman, previously a staff research scientist at Google DeepMind. Its founding team also includes Stanford robotic-learning researcher Chelsea Finn and University of California, Berkeley professor Sergey Levine, alongside former Stripe executive Lachy Groom. Their work connects robot-learning research with the practical demands of deploying machines in homes and businesses.
How Physical Intelligence’s robot foundation models work
Physical Intelligence develops robot foundation models, trained on varied tasks and machines to provide a starting point for particular applications. Its first generalist model, π0, pronounced “pi-zero”, was introduced in October 2024 and belongs to a category called vision-language-action models, or VLAs, which connect visual information and language instructions with robot actions.
This is necessary as recognising an object and manipulating it are different problems. Identifying a shirt in a camera image is only the beginning; completing the job requires a sequence of coordinated movements that responds to what happens to the fabric.
Physical Intelligence’s approach to this combines knowledge inherited from vision-language models with training on robot experience. The resulting system produces motor commands, connecting an instruction with physical execution. Its early demonstrations included folding clothing, assembling cardboard boxes and clearing tables across different robot platforms.
This allows a developer working on a particular application to start with capabilities learned across a broader range of physical experience.
How π0.7 combines skills to handle unfamiliar tasks
Its π0.7 model, announced in April 2026, also explored what researchers call compositional generalisation: combining learned skills to tackle unfamiliar problems.
Its training combined robot demonstrations, human videos and autonomous runs, with context describing how tasks were performed. Instructions and images can guide execution; an image showing a desired object position, for example, gives the robot a goal for its next step.
In another experiment, researchers coached a robot through loading a sweet potato into an air fryer using step-by-step language instructions. Training included limited examples of other air-fryer interactions, and the model combined that experience with guidance to perform the task, applying existing knowledge to an unfamiliar sequence of actions.
Pi API: adapting robot AI to new applications
The company also offers its latest models through the Pi application programming interface, or API, launched in beta in October 2026. Developers upload data, fine-tune models on infrastructure managed by Physical Intelligence and access hosted inference, the process of running a trained model. Fine-tuning adapts an existing model through additional training for the developer’s application.
Examples published with the launch include Actor’s autonomous wheel loader loading material into a truck at a commercial site and Almond’s dual-arm Axol robot deburring spacers, removing rough edges from components used in its assembly. According to Physical Intelligence, Almond produced the model behind its demonstration after four days of iteration with the API.
Developers supply their application knowledge and robot data, using unsuccessful attempts to identify which examples to collect next.
Robot AI in practice: laundry folding and warehouse packaging
In partner deployments, improvements are measured through the work robots complete and the help they need from operators.
Weave Robotics, which develops robots for household tasks, described commercial systems folding laundry in businesses around the San Francisco Bay Area. In results published by Physical Intelligence in February 2026, Weave reported that adding its data to π0.6 pre-training reduced missed grasp sequences by 42% and interventions by 50%, compared with the version trained without that data.
The missed-grasp measure counted sequences of at least two consecutive misses per laundry load. Fewer interventions mean fewer occasions when a remote operator must step in to help finish the job.
In the same publication, warehouse robotics company Ultra presented a full-shift recording of π0.6 packaging customer orders at a deployment site and reported improvements in throughput, instruction-following and recovery from difficult situations. Human operators intervene when required, generating further training examples. Both sets of results were supplied by the partners.
Physical Intelligence, Skild AI and FieldAI compared
Physical Intelligence is not alone in this field. Skild AI pursues a similar ambition with its Skild Brain, designed to control different robot bodies in applications spanning inspection, mobile manipulation and packaging. It also describes providing skills such as grasping, handovers and navigation through API calls.
FieldAI focuses on autonomous operations in complex environments, including construction and industrial sites. It says its architecture incorporates physics-based reasoning and uncertainty estimates so robots can respond to hazards by slowing down, rerouting or gathering more information, with the complete autonomy system running onboard without cloud connectivity.
These approaches overlap without being identical: Physical Intelligence’s examples emphasise adaptable manipulation skills and model training, while FieldAI foregrounds complete field missions. Working across different robot types is a shared goal rather than a capability unique to one company.
Building the intelligence other robot companies use
Physical Intelligence’s partnerships show how this division of labour works. Weave brings knowledge of garments and laundry operations, while Ultra understands packaging and order fulfilment; each contributes data from its own work to improve the capabilities it uses.
For Physical Intelligence, that creates a role across several branches of robotics. Specialist companies build machines and services around their customers, while its models give them a foundation to adapt. Laundry folding and warehouse packaging already illustrate how a shared research effort can support different businesses doing practical work.
Liked this article? You can support our independent journalism via our page on Buy Me a Coffee. It helps keep MoveTheNeedle.news focused on depth, not clicks.