The difficulty of humanoid robots lies not only in making them move or recognize objects, but also in building a computing and security system capable of handling a large number of sensors and immediate decisions while the machine operates near people and in homes, hospitals, and schools. According to an analysis published by Semiconductor Engineering on September 3, 2026, this proximity makes the risks of humanoid robots different from those of self-driving cars, because hacking a digital system in a robot could directly turn into a physical impact on people and the surrounding environment.
A Hybrid Architecture Between Centralization and Distribution
Humanoid robot designs are moving toward a hybrid model that combines a powerful central computing unit with distributed units located near the limbs and joints. The central unit handles general perception, scene understanding, planning, and robot-wide motion coordination, while smaller units process finger and joint data and local control loops with short response times.
This system can include microcontroller units (MCUs), general-purpose processors, graphics processing units, image- and signal-processing processors, as well as dedicated neural units to accelerate artificial intelligence. Ronald Stärz, a humanoid robotics systems engineer at Infineon Technologies, noted that a safety-dedicated controller should accompany the main processing system, and that safety should not be left to the artificial intelligence model alone.
Edo Cohen, head of the Physical AI Birds of a Feather group at MIPI Alliance, sees local computing as useful for fast loops involving the hands, feet, balance, and haptic feedback. Centralization, by contrast, may be more efficient in terms of component count, cost, and power consumption, while simplifying software development. Therefore, no single design suitable for all uses has emerged so far.
Sensors Determine the Shape of the System
Humanoid robots receive data from cameras, microphones, radar, lidar, and ultrasonic sensors, in addition to magnetic, capacitive, and resistive touch sensors. The fingers, in particular, need continuous feedback to adjust grip strength and prevent crushing an object or dropping it.
In a centralized design, relatively simple sensors can be used to send raw data to an FPGA-based sensor bridge, such as Nvidia HoloScan, and then transmit it over Ethernet to a GPU unit for scene analysis. This makes it possible to add a larger number of small sensors and shift a greater portion of power consumption to a single computing point, but the GPU unit itself may consume significant power.
Distributing processing near the sensors reduces response time and sends preprocessed data to the central unit, but it increases the number of components, software elements, and connection points. According to Nebu Philips of Synaptics, the decision depends on the entire processing pipeline, beginning with data reception, decoding, and image processing, through to the point where the data is aggregated, as well as on the type of robot, whether humanoid, collaborative, or service-oriented.
From the Edge to the Cloud
Choosing local computing does not necessarily mean isolating the robot from the cloud. Some systems may need to connect to data centers to train large language models or vision-language-action models (VLAs), while other applications benefit from small, specialized language models running on the device.
The choice depends on the company’s priorities: privacy, control over data, operating costs, the nature of the work environment, and the availability of wired or wireless connections to data centers. Matthew Bubis of Imagination Technologies explained that local and centralized computing will continue to coexist, since there is no single solution suitable for every company and task.
A Broader Attack Surface Than Cars
Humanoid robots combine dense sensor arrays, artificial intelligence models, wireless communications, and the ability to move and handle objects. Potential attack points therefore include model software, over-the-air firmware updates, camera drivers, Wi-Fi channels, control interfaces, and mechanisms for connecting intelligent agents.
Dana Neustadter of Synopsys warned of the risks of tampering with model integrity during loading or updating, as well as data-poisoning attacks through wireless updates. Exploiting the visual input could cause the robot to grasp the wrong object, enter a dangerous area, or perform an unsafe movement. Man-in-the-middle attacks or breaches of wireless communications could also enable control of the robot and the sending of commands to it.
Sylvain Guilley of Secure-IC believes that a human-like appearance may create additional risks, such as difficulty distinguishing between an automated agent and an authorized person, or using the robot for close-range eavesdropping or impersonating someone else. According to the premise presented in the article, voice-based and natural interactions could become a channel for exploiting trust, particularly when users interact with the robot as though it were a human party.
What Changes in Practice?
The most important conclusion for developers is that security and safety must be designed together with the architecture of the chips and communications, rather than added later as a software layer. Controls mentioned by the participants include an embedded device identity, secure boot, protected key storage, encrypted communications, and protection while models are running. The importance of multimodal authentication also stands out; voice alone may not be sufficient and can be supported by touch, a fingerprint, or a combination of camera and voice to verify identity and context.
The article also points to the importance of open standards, such as MIPI and RISC-V specifications, in providing suppliers with more options and shortening development cycles. However, openness does not eliminate the need for security verification and risk management, nor does it by itself settle the choice between a centralized and distributed design.
Power and cost remain decisive constraints. Excessive use of GPUs or sensors could shorten operating time or raise the price to a level that weakens commercial viability. In the United States, robot sales reached $11.4 billion in 2026, up 29% year over year, according to the Robotics Center report mentioned in the article. However, broader expansion of humanoid robots will remain tied to improving battery life, lowering costs, and demonstrating that systems can withstand hacking.
Editorial reading: The source does not announce a single platform or a complete security solution; rather, it depicts a design phase in which the architecture has not yet stabilized. The actual challenge is balancing response time, power consumption, the number of sensors, software flexibility, and safety and security requirements. The warnings concerning model manipulation and human interaction also reveal that protecting a robot does not stop at the chip or the network, but extends to the behavior of agents and the ways users rely on their outputs. The nature of regulatory requirements and the practical mechanisms for proving the safety of these systems remain open questions in the article, which conveys the views of experts from multiple companies and does not present a unified market standard.