Physical IntelligenceUnited States2026

π0.7

An AI model: a vision-language-action model (VLA) takes camera images and an instruction and turns them into movements for a robot. This page describes the software, not a robot.

Robotsnot stated

Key figures as stated by the developer

5 of 6 stated · 1 gap

Model type ⁠B vision-language-action
Controls ⁠B arms and hands
Inputs ⁠A camera language robot state
Access not stated
Parameters ⁠A approximately 5,000,000,000
Base model ⁠A Gemma3 4B

How to read the figures

14kgA30.9lbA: The letter in the grey square after a figure points to the source the figure came from. Robots with several sources carry several letters.
1,100mmB43.3inB: A dashed letter marks a secondary source – a datasheet, manual or developer documentation from the manufacturer, or a press report used only where the manufacturer gives no figure.
74.5lb: Imperial figures are our own conversion unless the manufacturer states an imperial figure itself. The source letter always belongs to the metric figure — a conversion has no source of its own.

The four states

Versions and notes in the key figures (5)

Model type

Developer wording: "VLA Vision-Language-Action Model".

Controls

"Control modality labels that indicate whether to use joint or end-effector control." The developer describes the model as one that "can perform a wide range of dexterous tasks".

Inputs

"The model takes as input up to four camera images (front view, two wrist views, and optionally rear view)". "We also feed the proprioceptive state" of the robot. Developer blog: "At test time, our model accepts standard language instructions, but also information about the desired strategy, and even synthetically generated visual subgoals produced by a lightweight world model."

Parameters

"The model has about 5B total parameters." Action expert: "a flow matching action expert with 860M parameters".

Base model

"The model consists of a VLM backbone initialized from the Gemma3 4B-parameter VLM".

Full specifications

8 of 12 fields stated

The letter points to the source at the foot of the page. Fields without a figure are dashed: the developer does not state them — that is not the same as zero, and not the same as no.
Field	Value	Source and caveats
Input and output
Model type	`vision-language-action`	B Developer wording: "VLA Vision-Language-Action Model".
Inputs	`camera` `language` `robot state`	A "The model takes as input up to four camera images (front view, two wrist views, and optionally rear view)". "We also feed the proprioceptive state" of the robot. Developer blog: "At test time, our model accepts standard language instructions, but also information about the desired strategy, and even synthetically generated visual subgoals produced by a lightweight world model."
Controls	`arms and hands`	B "Control modality labels that indicate whether to use joint or end-effector control." The developer describes the model as one that "can perform a wide range of dexterous tasks".
Degrees of freedom controlled	not stated
Size and speed
Parameters	approximately 5,000,000,000	A "The model has about 5B total parameters." Action expert: "a flow matching action expert with 860M parameters".
Control rate	50Hz	A Action execution layer, the fastest stated: "The UR5e robots run at 20 Hz, while all other robots run at 50 Hz." The paper gives a maximum inference latency of "240ms on a 50Hz robot".
Where it runs and what it is built on
Runs where	not stated
Base model	Gemma3 4B	A "The model consists of a VLM backbone initialized from the Gemma3 4B-parameter VLM".
Openness
Access	not stated
License	not stated
The developer’s own claims
Learns new tasks	without examples	B Developer claim: "it can follow new language commands and perform tasks that were never seen in its training data". For new kitchen appliances: "We did not collect demonstrations of these specific appliance tasks"; "For each appliance, the robot received language coaching for using the appliance".
Works across robots	yes	B Developer claim: "it can generalize across robot platforms, scenes, and tasks more effectively than our prior models"; for a bimanual UR5e system, "even though we did not collect any data of laundry folding with this robot". Yes only if the developer itself says the model runs on different robot models, for example robots from two manufacturers or two of one manufacturer’s models. Several units of the same robot do not count.

Notes

Country and city from the company registries CB Insights and PitchBook, which agree (396 Treat Avenue, San Francisco); the developer does not state its location.
Physical Intelligence describes π0.7 as a steerable generalist model that can perform dexterous tasks across robots, scenes, and skills.

Sources

A Secondary: arxiv.org/html/2604.15483 · Retrieved 2026-10-13
B Manufacturer product page. www.pi.website/blog/pi07 · Retrieved 2026-10-13

One line per source, not per figure: a source has one retrieval date. The letter beside a figure points to the line it came from.