🤖 Robonomy

DARPA's 'Physical Intelligence' RFI — Embedding Compute Into Robot Materials

3 min read

DARPA isn’t asking for better processors. It’s asking whether intelligence can live in the material itself.

On April 27, 2026, DARPA issued a request for information that reads like a philosophy paper with a budget line. The topic: “Materials for Physical Compute in Untethered Robotics.” The RFI (SN-26-76) closed May 27, 2026.

The core argument: current robots are assemblies of wire, metal frames, and motors that “yield a robot with small behavior diversity.” Even the most advanced systems require constant internal data processing with end-users or data centers, creating latency delays and consuming transmission power.

DARPA wants concepts at “the material, component, and kernel level” — down to chemistry and physics — that can change the nature of machine intelligence itself.

What “Physical Intelligence” Means

The RFI is asking for approaches where computation is intrinsic to the material, not bolted on as a separate compute module. Examples from the research community include:

  • Mechanical neural networks — Lattice structures that compute through physical deformation
  • Soft robotics with embedded logic — Pneumatic or hydraulic systems where the fluid pathways perform computation
  • Material state machines — Substances that transition between configurations based on environmental stimulus, effectively “processing” inputs through physical change
  • Oscillatory computing — Coupled mechanical or chemical oscillators that solve optimization problems through synchronization

The Untethered Constraint

The “untethered” part matters. Cloud-dependent robots fail when the link drops. Edge-dependent robots fail when battery runs low. DARPA wants systems that reason without either — computing through the physics of their own bodies.

This isn’t about efficiency. It’s about survivability in contested environments where electromagnetic denial is assumed.

The Companion Program: DICE

Alongside Physical Intelligence, DARPA is launching the “Decentralized Artificial Intelligence through Controlled Emergence” (DICE) program, led by Program Manager Susmit Jha in the Information Innovation Office. The program was announced via Special Notice DARPA-SN-26-65 on April 21, 2026, with a Proposers Day held May 29, 2026 (registration via SN-26-72).

DICE targets swarm coordination — machines that “dynamically form teams using peer-to-peer coordination to execute complex missions” without a central controller. The program explicitly excludes real-world autonomous system deployment; it will use simulation environments to demonstrate architectures in defense-relevant use cases.

Together, the two efforts map a future where individual robots are smarter at the material level and swarm-level intelligence emerges from local interaction, not top-down command.

Why This Is Foundational

Most autonomy research treats the robot body as fixed and the software as variable. DARPA is asking whether the body itself can be the computer.

If this research succeeds, the implications extend far beyond military systems:

  • Humanoid robots that don’t need GPUs consuming 300W of battery
  • Swarms that coordinate through physical coupling rather than radio links
  • Wearable robotics where the fabric computes

The Timeline

The Physical Intelligence RFI (SN-26-76) closed May 27, 2026. DARPA typically uses RFIs to scope formal program solicitations, which would follow 6–12 months later. A funded program would run 3–5 years.

DICE is at the Proposers Day stage (May 29, 2026); a formal solicitation would follow if DARPA proceeds.

This is early-stage, high-risk research. Not a product announcement. But DARPA’s historical pattern is to invest in foundational shifts 10–15 years before they reach fielded systems.

Sources: DARPA official (darpa.mil), SAM.gov, Defense One, EverGlade | Last updated: 2026-05-27

Graph View

Recent