Article reader Listen + reading controls
Article reader
Preparing the reader…
Reading settings
Autonomy is also a materials problem¶
The Defense Advanced Research Projects Agency (DARPA) is asking researchers to rethink robotics through physical intelligence: materials and structures that integrate sensing, adaptation, computation, and actuation rather than sending every signal through a centralized processor.
The idea is technically ambitious. It is also a useful corrective to the way artificial intelligence conversations often collapse an entire system into its software.
Where intelligence lives shapes resilience¶
A robot operating in a laboratory can rely on stable power, high-bandwidth communication, carefully maintained sensors, and abundant compute. A defense system may face distance, interference, damage, heat, dust, limited energy, and an adversary actively trying to break the loop.
Under those conditions, architecture becomes strategy. A system that must transmit every observation to a central model inherits communication delay and a single point of dependence. A structure that can respond mechanically or locally may continue functioning when computation or connectivity is degraded.
Biological systems offer the broad analogy: useful behavior is distributed across morphology, reflex, sensing, and central cognition. The body does not ask the brain to calculate every correction required to remain upright.
Capability is allocated across the whole system¶
Systems engineering has long emphasized that a system's behavior emerges from interactions among components. Physical intelligence makes those allocations visible. A compliant material can reduce the control precision required for grasping. A sensor-actuator structure can react before a remote controller receives the data. Mechanical design can make unsafe states physically difficult to reach.
That can simplify some software problems while creating new verification challenges. A material may change with temperature, fatigue, manufacturing variation, or damage. Its “decision” may not produce a conventional log. Inspecting why a structure responded as it did may require different measurement tools and new expertise.
The International Council on Systems Engineering's systems engineering handbook is relevant here because requirements, interfaces, verification, and lifecycle concerns must cross disciplinary boundaries. Materials scientists, roboticists, control engineers, artificial intelligence (AI) researchers, safety specialists, manufacturers, maintainers, and operators need a shared model of the capability.
Design degradation, not just peak performance¶
Mission systems are rarely binary. They degrade. Trustworthy autonomy needs to make that degradation predictable and legible.
Teams exploring physical intelligence should ask:
- Which functions remain available when communication or central compute is lost?
- How do material aging and environmental conditions change response?
- Can the system detect when local behavior is outside its qualified range?
- Which failures are graceful, and which cascade?
- What can an operator or maintainer observe without specialized laboratory equipment?
- How will configuration and manufacturing variation be controlled?
These questions turn an intriguing material into a supportable capability. They also create the evidence needed to decide where physical response should end and higher-level decision-making should begin.
Organizational boundaries must move with the architecture¶
When intelligence is distributed into hardware, responsibility cannot remain organized as though software alone owns autonomy. Test plans, safety cases, maintenance practices, and change boards must reflect the new allocation. A material substitution may now be an algorithmic change. A repair procedure may alter control behavior. A supplier-quality issue may become a mission-assurance issue.
DARPA's request for information opens a promising research direction for systems expected to operate at the edge. Its larger lesson is already usable: robust autonomy comes from designing the complete loop under real constraints.
The model matters. So do the sensor, structure, actuator, power budget, communication path, human supervisor, and maintenance system. Intelligence is not simply added to a machine. It is allocated across one.
Sources and research trail¶
- Defense Advanced Research Projects Agency, “Rethinking Robotics with Physical Intelligence” (April 29, 2026).
- International Council on Systems Engineering, Systems Engineering Handbook, Fifth Edition (2023).
- Simon, The Sciences of the Artificial (1996).
- Pfeifer and Bongard, How the Body Shapes the Way We Think (2006).
- National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0) (2023).