[Tech Report] Physical Intelligence That Boosts Humanoid Reliability

Humanoid robots are moving out of controlled demo environments. They are gradually expanding into factories, warehouses, medical settings, and other spaces shared with humans.

Success in these environments depends on the ability to operate continuously and reliably amid real-world uncertainty. Mobility is improving accessibility. The ability to perform delicate tasks is expanding use cases.

Wider adoption and use of humanoids will hinge on data and reliability. Data underpins AI models that extend humanoid autonomy. Data underpins the norms robots learn. Data underpins sensor fusion systems.

Reliability is tied to the system's ability to ensure stable operation in the physical world. 'Physical intelligence(physical intelligence: PI)' plays that role.

'Physical intelligence(physical intelligence: PI)' includes continuous sensing. It also includes deterministic connectivity, real-time control, and reactions based on physical laws. It refers to AI inference that operates within physical constraints.

Humanoid robotics marks a turning point. The background to this turning point is not the independent advancement of intelligence.

The scope of innovation spans the entire physical control loop, including sensing, consistent connectivity, computing, motion, power, simulation, and safety. These innovations progress in a converged manner.

The advanced humanoid era cannot be realized by AI alone. It also cannot be realized by hardware alone.

The conditions for implementation are close interaction across each domain. The premise for such interaction is physical-signal-based operation and compliance with real-time constraints.

These combined effects drive changes in robots. The direction of change is a shift from rigid automation systems to systems equipped with physical intelligence, as shown in Figure 1.

Conventional industrial robots were designed around predictability, operated within structured areas, and executed preprogrammed routines. Safety fences, light curtains, and strict zoning were used as safety measures, and isolation was the risk-mitigation approach.

Humanoid robots are in a position that goes beyond existing models. Their operating environments are open and dynamic, and they move while sharing space with people. That makes it necessary to respond continuously to unexpected situations.

In these conditions, intelligence must be continuous. It also needs to comply with limited time constraints and must have a physical implementation. This is not merely execution at the preprogramming stage; it leads to the emergence of a new system type grounded in physical intelligence.

The operating elements of this system are sensing the situation, deciding how to respond, acting in response to the environment, observing the outcome, and making immediate adjustments. In physical systems, there is no pause between steps. Time does not stop, and intelligence must act with precise timing.

In the physical world, this loop becomes the minimum requirement for stability, safety, and useful work. If that connection breaks down, delays, jitter, and missed updates appear as hesitation, trembling, and abrupt stops. Therefore, maintaining an uninterrupted continuous loop of sensing, judgment, action, observation, and adjustment becomes the basic condition for safe and useful physical work.

When the loop is properly coordinated, the robot shows smooth movement. It can also respond effectively to uncertainty and collaborate naturally with people.

Conversely, if this loop collapses, the impact is far from minor. The connection state of the coordinated loop affects the robot's movement, response, and collaboration as a whole.

For this reason, the essence of physical intelligence is presented as a systems engineering problem. It requires close alignment among AI models, algorithms, perception, decision-making, and control.

When these conditions are met, it becomes possible to implement a physical intelligence system based on precise coordination and temporal synchronization. The discussion then shifts to Multimodal Perception.

Humans do not rely on a single sense. They continuously and jointly use vision, hearing, touch, balance, proprioception, and force feedback to understand their surroundings.

Humanoids in real environments also need to perceive their surroundings in the same way humans do. There is also an additional set of conditions for humanoids operating in real environments.

Humanoids need structured cognitive capabilities that can understand and reason about the physical world beyond AI scene classification. They operate based on continuous information flow across sensors, computing nodes, and actuators, and use that to perceive, interpret, and respond at the right moment. The related chart is in Figure 2.

A recent trend in humanoids is the expansion of fused sensing information. Vision is responsible for spatial understanding and grasping semantic context, while depth sensing handles terrain and distance. Audio and vibration are used for intent recognition, environmental context recognition, and awareness of out-of-view situations.

Tactile sensing contributes to object manipulation, contact detection, and grip stability. IMU and joint encoders are used to determine balance, orientation, and proprioception. Force and torque sensing are used for safe interaction, load estimation, and application of physical constraints.

The interconnection of multiple sensing methods strengthens perception performance beyond what individual sensors can achieve. This makes robust perception possible even amid motion, glare, occlusion, complex environments, noise, and uncertainty.

Such perception is a prerequisite for a humanoid's intelligent behavior. The requirements for perception are continuity and situational appropriateness.

The role of perception is to generate data. The role of connectivity is to move and synchronize that data.

Humanoid architectures are moving from single-structure machines to more distributed, interconnected networks. The decisive factor in multimodal data insight is what each sensor measures and when it measures it. If temporal consistency is maintained across sensor signals, the scope of AI reasoning expands to cause, contact, movement, and outcome.

The connectivity among camera areas, sensing areas, and actuation areas is hierarchical, and the related diagram is mentioned in Figure 3. In this hierarchical link, edge bus throughput provides a midrange data rate of 1 Mb to 10 Mb, while high-speed backbone throughput is 2 Gb to 10 Gb.

For distributed intelligence in robots to function properly, consistent connectivity must be in place. Without connectivity support, a distributed intelligence system cannot operate normally. Conversely, when consistent connectivity is secured, the robot can have an integrated sense of itself and its surroundings.

Based on this integrated sense, the robot can achieve smoother movement. Safer collaboration also becomes possible. In addition, adaptive manipulation becomes possible, along with consistent behavior even in uncertain situations.

In 'Thinking, Fast and Slow,' Daniel Kahneman distinguished between fast, intuitive responses and slow, deliberative reasoning. This distinction is presented as System 1 and System 2. When applied to humanoids, this concept requires a physical foundation that can actually function.

That physical foundation is System 0, a real-time neural system composed of sensors. The performance of humanoids depends on systems that operate within a single synchronized loop of touch, thinking, and torque. Ultimately, to implement the distinction between System 1 and System 2 in humanoids, System 0 must be the premise.

In a crowded industrial site, a humanoid carrying a not-fully-filled box with the lid open through an aisle may seem simple, but it is actually a task involving multiple system domains at the same time. This task requires several system layers with different time scales and roles to work together.

System 0 at the lowest level handles the physical-signal and actuation base and operates on a time scale of less than 5 ms. This layer consists of sensors, actuators, joint states, motor current, tactile signals, IMU updates, and low-latency local loops. Functionally, it means securing the robot's actual physical signal base.

System 1 above it handles fast interpretation and reflexive response and operates on a time scale of 5 ms to 20 ms. This layer performs sensor fusion, slip response, balance correction, force adjustment, local motion control, contact modeling, and safety enhancement. Its functional meaning is to turn signals into fast, corrected actions.

System 2 handles cognitive reasoning and task planning and operates on a time scale of 50 ms or more. This layer is responsible for path planning, human motion prediction, human intent prediction, task-goal understanding, norm-level decision-making, and planning. Functionally, it means reasoning about task goals and environmental context.

Physical intelligence emerges when coordination across these layers is achieved. The fastest physical-signal and actuation layer provides the foundation, fast interpretation and response above it correct the action, and the slower cognitive and planning layers handle goals and surrounding context. Ultimately, for a robot to operate successfully, synchronization among sensing, reasoning, and action must be maintained.

Some evaluation criteria for humanoids include motion completeness. Examples of motion completeness include stable walking, adaptive arm movement during motion, and balance correction without noticeable hesitation.

These motions are achieved through coordinated control across dozens of DoF, as shown in Figure 4. The natural and stable movement of humanoids requires control across multiple DoF to mesh together.

As the number of DoF increases, functionality expands. But as DoF increase, control complexity also increases.

Each time a joint is added, coupling increases, timing margin decreases, and instability paths are added. As a result, even small timing errors spread quickly.

If ankle correction is delayed, knee load changes; if wrist disturbance occurs, shoulder torque changes; and if the torso is adjusted, full-body balance must be regained. In situations where effects cascade across other joints and the whole body balance, real-time motor control becomes the core foundation.

Therefore, controlling many DoF requires continuously understanding joint states across position, motion, and load, and doing so reliably. All of this must be maintained as an integrated process of real-time motor control and continuous, reliable understanding of joint states, even during power outages, dynamic load conditions, and mechanical wear.

Rotary actuator types use gear reduction, while linear actuator drive methods use lead screws or ball screws. However, the quality of robot motion depends, regardless of actuator type, on high-precision position sensing, low-noise signal chains, and actuator-level predictable timing.

Designing sensing, motor drive, power, and control as a single coherent system makes it possible to expand robot motion smoothly. This expansion occurs under the condition that there are no failures that introduce timing uncertainty.

Complexity does not disappear in this process. What becomes possible is consistent management of complexity.

For digital intent to be implemented as physical motion, signal integrity and timing must be maintained at every joint. Signal integrity and timing at every cycle, as well as in every physical interaction, are also presented as implementation conditions.

Along with this, the topic of Functional Safety is raised. The innovativeness of humanoids is recognized when collaboration close to humans becomes possible, while conventional industrial automation environments have operated with safety-fence separation. But as the environment shifts toward humans and robots sharing the same space, the concept of safety has had to be redefined.

Mechanical compliance and back-drivability are useful for reducing the risk of injury. However, those characteristics alone are not a sufficient condition for safety. Even if the mechanical motion is smooth, safety is not guaranteed if sensing errors occur, and safety is not guaranteed if timing is inconsistent. Even if the mechanical motion is smooth, safety is not guaranteed if an unexpected failure occurs.

For this reason, Functional Safety in human-proximate situations requires layered protective measures. The requirements include setting limits on force, torque, speed, and work space. They also include continuous diagnostics, sensor calibration, plausibility checks, sensor redundancy, diverse safety paths, and consistent responses under failure conditions. In addition, intentionally controlled performance degradation rather than unexpected failure is included among the Functional Safety requirements.

Because humanoids are interconnected systems, safety cannot be limited to a single module. Safety applies across perception, connectivity, computing, power, and actuation. Therefore, safety must be applied from start to finish.

The key to wider robot adoption depends on whether the robot behaves predictably and consistently when a failure occurs. This is the deciding factor in whether robot deployment will expand or remain limited to impressive demos.

This criterion is also linked to the conditions for building trust. Trust is formed not by perfect operation itself, but by whether safe, consistent, and transparent responses are made in unexpected situations.

One example that most compactly reflects these demands is the 'humanoid hand' realized through 'AI, sensing, and control' and physical intelligence. The humanoid hand is the area where system requirements are most concentrated.

In the hand, sensing, actuation, wiring, and reflex-level control are required as converged elements. At the same time, the hand's constraints are among the strictest in terms of latency, power, heat, and size.

Human hands adjust gripping force almost unconsciously and detect slipping before an object falls. They also adjust contact pressure according to surface differences and reposition objects with fine corrections while moving, and this repositioning is done with micro-adjustments so subtle that they are barely consciously perceived.

To implement some capabilities of a humanoid hand, advances in high-density tactile sensing, precision motion control, miniaturized actuation, low-latency local computing, multimodal sensing-based learning policies, and robust power management are prerequisites.

All of these capabilities must be safe to use around people and must be implemented within a lightweight form factor and a highly durable form factor.

Delicate manipulation depends on comprehensively understanding reality through various kinds of sensing. Vision sensing identifies the type of object and its position, while depth sensing identifies the object's shape and distance. Tactile sensing judges whether there is slippage and whether force is appropriate, and proprioception identifies task-related joint states and body states. Force and torque sensing determine whether the interaction force is within a safe range and a useful range, while audio and vibration sensing detect contact signals, mechanical anomaly signals, and task-completion signals.

In discussions of robot perception, vision and depth sensing tend to be emphasized. By contrast, audio tends to be underestimated. But audio has practical added value. Mechanical anomalies can be detected based on the vibration and acoustic characteristics of sound. A click sound is also used to verify whether a connector has been fully inserted and secured. In addition, characteristic sound signals can be used to confirm whether an action has been completed.

Tactile sensing then provides unique information that is difficult to obtain through vision and audio. That information relates to changes in contact-surface interaction. Tactile sensing detects pressure distribution. It also detects vibration and changes in friction. Furthermore, fine slip for judging grip stability is included among the detection targets.

This shows that the essence of the hand lies in refined dexterity. Refined dexterity is a mechanical challenge. At the same time, it is also a system-level challenge that combines sensing, computing, control, and AI.

The related diagram is in Figure 5.

As humanoid robots increasingly enter human environments, the importance of hand intelligence is also rising, and it is being mentioned on par with motor force. The idea is that as humanoids move into human work and living environments, the role of the hand becomes as important as driving force.

A complete implementation of the hand is presented as the key to responding to the many changes in real industrial environments. At the same time, complete implementation of the hand is also cited as the element that turns humanoids into true general-purpose machines.

The humanoid hand is a space where sensing, computing, and intelligence converge with every grasping action. For this reason, the most difficult challenge in humanoid robotics is said not to be limited to individual component issues.

The points where problems arise are also said to be concentrated more at system interfaces than in components alone. Specifically, the sensing-computing interface, the control-power interface, and the connectivity-safety interface are presented as points where problems occur.

For this reason, physical intelligence is described not as a simple software layer added on top of robot hardware. Rather than being separate software, it is presented as an integrated characteristic that determines field performance.

Accordingly, two humanoids using similar sensors, motors, and AI models may still differ in real-world performance. This shows that, in the process of combining multiple functions around the hand, integrated characteristics can determine performance.

The success or failure of humanoids depends on systems integration rather than individual functions. The core criteria for systems integration are maintaining timing amid real-world changes, maintaining signal integrity amid real-world changes, maintaining power management amid real-world changes, and maintaining safety amid real-world changes.

Analog Devices says it is approaching these challenges based on decades of experience in precision sensing, decades of experience in consistent connectivity, decades of experience in real-time signal processing, decades of experience in power management, decades of experience in isolation, and decades of experience in Functional Safety.

Analog Devices said it provides system-level understanding of the most demanding challenges in humanoid design. It also defined its role not as a simple technology supplier, but as a partner that supports predictable operation in real-world complex physical systems.

Wider adoption of humanoid robots will not be achieved by graceful walking alone, and it will not be achieved by object manipulation in controlled environments alone. Humanoid proliferation can happen only when continuous operation, uncertainty adaptation, failure recovery, and safe behavior around people are possible.

These conditions connect to the realm of physical intelligence. Physical intelligence is defined as AI inference tuned to physical constraints.

The future being presented will be realized not by handling intelligence, perception, motion, connectivity, power, and safety separately, but by co-designing them as a single synchronized system. The key method for this is consistent systems engineering.

Reliable physical intelligence is not formed by chance; it is formed non-accidentally. In the implementation process, a consistent control loop must be built step by step.

As robotic systems evolve, the challenge also changes. The focus shifts from adding functions to organically integrating and coordinating functions.

Source: TECHWORLD · Tracy Johnson, ADI Industrial Automation Product Marketing Director
Original: https://www.epnc.co.kr/news/articleView.html?idxno=407706

References

• https://www.epnc.co.kr/news/articleView.html?idxno=407706