By Dara O'Sullivan, Tracey Johnson & Margaret Naughton
As robots move into open environments where they interact with people, determining where sensing, decision making, and control run becomes critical for stability, safety, responsiveness, and trust. Humanoids face a fundamental whole-body control problem: every joint, sensor, and actuator must coordinate to maintain balance and purposeful motion. If coordination breaks down, the robot falls.
The previous blog defined physical intelligence and why intelligence changes when it leaves the abstract world of data and enters the physical world of motion, force, and interaction. This blog explores where robotic intelligence actually lives and how responsibilities are split across central compute, edge systems, and local control and why these architectural choices matter.
Why Real Time Matters: The World Doesn’t Wait
Modern robots behave as distributed systems, where perception, state estimation, policy, planning, and control are continuously coordinated across the body to produce stable, real-time behavior.
In physical systems, latency is a stability problem: the difference between a stable step and a fall, a controlled grasp and a dropped part, a smooth stop and a safety incident. For behaviors like balance correction, collision reflexes, slip detection, and contact-aware motion, tens of milliseconds matter. Jitter, or unpredictable timing, can be worse than a fixed delay.
This leads to a first principle of physical intelligence: if the robot moves, the intelligence governing that motion can be centralized, but the control of motion must happen locally and on a predictable and deterministic schedule.
Speed Alone Is Not Enough: Determinism Is the Foundation
Robots need deterministic latency, not just low latency. Physical intelligence operates as a continuous loop: sensing, deciding, acting, observing, and adjusting. That loop must execute on time, every time. When timing slips, the loop destabilizes, causing oscillation, hesitation, collisions, dropped objects, or hard stops.
Deterministic operation is the substrate everything else depends on. Without it, safety mechanisms fail and human collaboration becomes impossible. With it, robots can move smoothly and remain stable in the face of uncertainty.
Whole-Body Intelligence: Central Policy, Distributed Execution
Humanoids intensify the architectural tension between centralization and distribution. The robot’s entire body is coupled, benefiting from a central “brain” that can reason over the whole-body state and coordinates motion as one coherent system. This drives designs toward highly centralized intelligence.
Central compute handles global state estimation, whole-body coordination, learned policy inference, task-level reasoning, intent, and long-horizon planning. These functions benefit from global context and a unified view of the robot.
However, if every tactile event, motor current spike, joint encoder tick, inertial measurement unit (IMU) update, and camera stream ships to that central brain in raw form, the robot quickly becomes a bandwidth and scheduling nightmare. Edge nodes must perform data prioritization and filtering, deciding what to send, what not to send, and what is urgent. The pattern is centralized coordination, distributed execution, and embedded reflexes, all synchronized in time.
Figure 1: The Distributed Brain of Modern Industrial Systems
Three Layers of Intelligence (and Why Each Exists)
- Central Intelligence: Intent, Policy, and Whole-Body Coordination
This layer knows what the robot is trying to do. Central intelligence handles global coordination, but rarely runs the fastest reflexes. In humanoids, this matters because coordination is holistic: controlling an ankle, shoulder, and torso is not separable when balance is at stake. Modern systems increasingly run AI-based policy learning here for complex, high-dimensional control.
- Edge Intelligence: Real-Time Control, Safety, and Data Reduction
This layer maintains system responsiveness under real-world timing constraints. It stabilizes motion, enforces safe torque and force limits, manages local sensor fusion, and reduces high-rate signals into decision-critical information. Instead of streaming everything upstream, the system filters and aggregates locally.
- Embedded Intelligence: Reflexes at the Actuator and Contact Point
The fastest control loops live here. Every actuator has its own local motor controller, responding to events that unfold too quickly to tolerate additional latency. This reflex layer includes motor current regulation, joint position control, safety enforcement, and the detection of micro-scale contact events like tactile spikes and slip events. These loops don’t call the main computer: the safest response never leaves the joint, hand, or end effector.
A Simple Example: Carrying an Open Container
Consider a humanoid carrying a partially filled open container down a busy industrial aisle. What appears simple spans multiple time domains.
- Embedded Control and Reflexes (sub 5ms class): joint stabilization, local torque loops and limits, grip micro-adjustments, slip detection, and fast disturbance rejection. These loops must be local at the actuator/hand.
- Whole-Body Control (5–20ms class): whole-body planning, perception and motion sensor fusion, contact state estimation, joint coordination, obstacle proximity reactions, and safety enforcement.
- Learned Policy (50ms+): route selection, human motion prediction, task intent (“keep container upright while making progress”), and policy-level decisions.
Too much centralization produces slow, hesitant behavior. Too much decentralization produces fast reactions without coherent intent. Physical intelligence lives in the balance: global intent and local reflex, operating as one synchronized system.
Why This Architecture Matters
The distributed brain is not just a performance optimization. It is the foundation for:
- natural, stable motion
- safe collaboration with people
- dexterous manipulation and contact awareness
- real-world operation under uncertainty
- resilient, reconfigurable industrial workflows
This is how robots move beyond automation and toward physical intelligence by putting intelligence where timing, interaction, and safety demand it.
Read all the blogs in the Humanoid Robotics series.