π0.7 (pi-zero-point-seven)

What it is

π0.7 is a robot learning model published by Physical Intelligence on 16 April 2026 ^[1]. The company calls it a general-purpose model: instead of being built for a single task, it performs a wide range of dexterous manipulation tasks at the same performance as fine-tuned specialists ^[1]. A "specialist" here means a policy that has been trained further on data from one narrow task; the claim is that a single general model can now stand in for a collection of such task-specific models.

How it was trained

The training data came from several very different places: many different robots, human data, and autonomous episodes collected by running various policies ^[1]. Mixing sources like these is awkward, because they differ in the physical machine, the sensors, the control interface and the quality of the behaviour. Physical Intelligence says the key to using them together was adding diverse context to the prompt, describing not only what the robot should do but how ^[1].

That context can take several forms. Prompts can carry language sub-steps that break a task into smaller instructions, metadata such as the desired speed and quality of execution, control-modality labels that say which kind of control the data represents, and visual subgoal images ^[1]. The effect is that the model is told the circumstances of each episode rather than being asked to average over them, so differences between data sources become information the model can condition on.

How it is used

At test time the model accepts a language instruction, a desired strategy and synthetic visual subgoals produced by a lightweight world model ^[1]. A world model is a component that predicts how a scene will look, so it can generate an image of a plausible intermediate state; those generated images then serve as targets the robot can aim for along the way. Because the strategy and the subgoals are inputs, a user can steer how a task is carried out, not just which task is carried out, without retraining the model.

What it can do

According to the company, π0.7 can follow new language commands and perform tasks that were never in its training data ^[1]. It also reports the first signs of compositional generalisation, meaning the model recombines skills it already has into new combinations: it used new kitchen appliances, and it folded laundry on a new robot for which there is no laundry-folding data ^[1].

The evaluation was deliberately demanding. π0.7 was tested on a bimanual UR5e system with Robotiq parallel jaw grippers, one of the least represented embodiments in its training set ^[1] — a two-armed setup unlike most of what the model had seen. The human demonstrations behind its data came from teleoperators with a mean of 375 hours of teleoperation experience ^[1], which is relevant because the quality of the demonstration data sets a ceiling on what the model can learn from it.

Where it sits

The author list includes Chelsea Finn, Sergey Levine, Karol Hausman, Brian Ichter and Suraj Nair ^[1]. The result matters because the usual trade-off in robot learning is between general models that are mediocre at everything and specialists that are good at one thing and must be rebuilt for the next; π0.7 is presented as closing that gap while remaining steerable through language and visual subgoals ^[1]. Compositional generalisation is the part with the widest implications: if a model can put existing skills together for a new appliance or a new robot without new data, then covering every task and every machine with demonstrations becomes less necessary ^[1].