Skip to content
Layer 04 · Embodied AI infrastructure

Physical Intelligence

The robotics and embodiment layer — humanoids, manipulators, and autonomous manufacturing systems that close the loop from decision to physical effect.

Role in the stack

What this layer does

The physical-intelligence layer closes the loop from coordination decision to physical effect. Without embodiment, planetary coordination is a planning exercise. With it, decisions become motion in the world.

This layer depends on the agentic layer for direction, the AI layer for reasoning, and the cognitive layer for inference primitives. The economic-orchestration layer depends on it: every supply-chain instruction eventually runs through manipulators, humanoids, and autonomous logistics.

Capabilities

What it provides

General-purpose manipulation

Dexterous manipulation across the kinds of objects and tasks that human-scale environments require. We treat this as a foundation-model problem, not a per-task engineering problem.

Whole-body humanoid control

Locomotion and whole-body coordination at human speeds in human-scaled environments. Hardware-software co-design with safe-by-construction torque limits.

Autonomous manufacturing primitives

Modular cells that compose into larger production lines, with reconfiguration measured in days rather than months.

Architecture

How it is composed

  1. 01Sensorimotor foundation models trained across multiple embodiments.
  2. 02Action-tokenization layer enabling cross-embodiment policy transfer.
  3. 03Real-time inference at the edge with cryostat-equivalent latency budgets in robotics.
  4. 04Aegis envelopes specifying physical safety invariants (torque limits, exclusion zones, contact constraints).
  5. 05Operator console with explicit physical-handoff and interrupt primitives.
Open challenges

What's hard

  1. 01

    Contact-rich manipulation at scale

    Most progress to date is on quasi-static manipulation. Dynamic, contact-rich tasks remain hard outside of narrow domains.

  2. 02

    Real-time inference budgets

    Closed-loop physical control requires inference within strict latency bounds. Foundation-model architectures and edge compute must co-evolve.

  3. 03

    Safety in shared spaces

    When humanoids work alongside humans, the safety case must be auditable and provable, not just empirical.

Roadmap

Where the work stands

  1. Shipped2025

    Sensorimotor research stack

    Internal pipeline for cross-embodiment policy training and evaluation.

  2. In progress2026

    Bench-top humanoid prototype

    Indoor manipulation and locomotion development platform.

  3. Planned2027+

    External operational pilots

    Bounded industrial pilots in well-characterized environments.

Related across the site