Skip to content

RoboticsPhysical Intelligence

Physical Intelligence

General-purpose foundation models that aim to control any robot to do any task.

Category
Robotics
Pricing
FREE
Hosting
Local
Platforms
APILinux
Models
Self-contained (on-device)
Verified
Jun 9, 2026

Physical Intelligence is a San Francisco lab building general-purpose AI for the physical world. Its π0 (pi-zero) vision-language-action model, built on a pretrained VLM and trained across many robot embodiments, can perform dexterous tasks like folding laundry. The company open-sourced π0's code and weights via its openpi repository.

Capabilities 5

What it actually does — grouped by capability family.

  • Fine-tuning / training (secondary capability)
  • Robot foundation model (primary capability)
  • Vision-language-action (secondary capability)
  • Dexterous manipulation (secondary capability)
  • Imitation learning (secondary capability)

Pros & cons

  • openpi repo runs and fine-tunes locally
  • Cross-embodiment VLA foundation model
  • Built on a pretrained vision-language model
  • Demonstrated dexterous real-world tasks
  • Research artifact, not a turnkey product
  • Needs robot hardware and ML expertise
  • Linux/local deployment only

Tags

Further reading

View all Robotics
  • View Skild AI details
    RoboticsPAID

    Skild AI

    Skild AI

    A single, omni-bodied foundation model designed to control any robot.

    Skild AI builds Skild Brain, a general-purpose robotics foundation model intended to control diverse embodiments — quadrupeds, humanoids, tabletop arms, mobile manipulators — without prior knowledge of a robot's exact form. It pretrains on large-scale simulation and internet video, then post-trains on targeted real-world data for customer deployments.

    Single model targets many embodiments
    Foundation model, not a shipping consumer product
    • robot-foundation-model
    • embodied-ai
    • humanoid
    • manipulation
  • View 1X details
    RoboticsPAID

    1X

    1X Technologies

    NEO, a home humanoid robot driven by the onboard Redwood vision-language-action model.

    1X (1X Technologies) builds NEO, a household humanoid robot powered by Redwood, a 160M-parameter vision-language-action model that jointly controls locomotion and manipulation for tasks like laundry and answering the door. Redwood runs on NEO's onboard GPU and pairs with an off-board language model for voice control. NEO ships from 2026 at $20,000 or $499/month.

    Redwood VLA runs onboard on NEO's GPU
    Ships from 2026 — unproven in real homes
    • humanoid
    • home-robot
    • vla
    • redwood
  • View Figure AI details
    RoboticsPAID

    Figure AI

    Figure AI

    Humanoid robots powered by Helix, an in-house vision-language-action model.

    Figure AI builds general-purpose humanoid robots (Figure 01–03) driven by Helix, its proprietary vision-language-action model that fuses perception, language, and learned control. Figure 03, announced in 2025, targets the home and broader applications; earlier models were piloted in BMW's automotive manufacturing.

    Proprietary Helix VLA model
    Not a buyable product yet
    • humanoid
    • vla
    • embodied-ai
    • helix
  • View Wayve details
    RoboticsPAID

    Wayve

    Wayve

    End-to-end embodied AI for self-driving that scales across any vehicle.

    Wayve builds AV2.0, an embodied-AI approach to automated driving that replaces hand-engineered perception-planning-control pipelines with a single end-to-end neural network mapping raw camera and radar input to driving commands. It develops the GAIA generative world model for training and the LINGO language model for interpretable driving. Backed by NVIDIA, Microsoft, and SoftBank, and partnered with Uber.

    End-to-end learned driving (AV2.0)
    Not a buyable product
    • autonomous-driving
    • embodied-ai
    • world-model
    • av2.0