How NXP Scales Physical AI from Drones to Humanoids

A delivery drone and a humanoid robot face the same core challenge: they need to reason, coordinate, and react at three distinct levels simultaneously. NXP Semiconductors’ Neural Access Architecture splits robotic intelligence into cerebral, cerebellum, and reflexive processing layers, each mapped to specific silicon. Presented at NXP Tech Days 2026 in Santa Clara by Ravi Annavajjhala, Vice President and General Manager, Neural Processing Unit (NPU), AI/Physical AI, and Altaf Hussain, Senior Robotics Segment Marketing Manager, NXP Semiconductors, the architecture scales from drones, where NXP covers approximately 100% of semiconductor content, to humanoid robots, where volume production is expected from 2032 onward. NXP currently covers roughly 50% of the semiconductor content in humanoid platforms.

How Does a Three-Layer Architecture Distribute Robotic Intelligence?

The Neural Access Architecture mirrors how biological nervous systems work. The cerebral layer runs reasoning, path planning, and foundational AI models including vision-language models. The cerebellum layer handles what needs to happen in real time: motion planning, sensor fusion, and safety management. The reflexive layer sits closest to the physical world, operating motor control loops, local actuation decisions, and a real-time safety network.

The same layered structure applies whether the robot is a delivery drone, an autonomous mobile robot (AMR), or a humanoid. For drones, NXP reuses existing technology and covers the full semiconductor bill of materials. AMRs are in near-term commercial deployment, with sensing that includes vision-only VSLAM, LiDAR combined with vision, and emerging indoor radar. Humanoid robots represent the largest long-term semiconductor content opportunity, though only tens to hundreds of units are deployed today.

“AI is at the edge, by default, across a diverse set of devices. Decisions are made on the devices themselves, and the edge is going to be something critical,” NXP’s NPU lead said during the session. Processing stays distributed across the robot body rather than bottlenecked in a single compute module.

What Connectivity Stack Do Robots Need Beyond TSN?

A three-layer brain needs a communication backbone to match. Gigabit Ethernet with Time-Sensitive Networking (TSN) is displacing EtherCAT as the dominant robotics protocol, bringing lower latency and the cost advantage that comes from automotive-scale Ethernet production.

Below TSN, three technologies handle specific parts of the robot body. Evolinks, an ASA-based sensor aggregation framework, replaces proprietary camera link standards like GMSL with an automotive-grade alternative. IEEE standardization for Evolinks is in progress, and NVIDIA’s Holoscan platform already streams sensor data through it, connecting perception hardware to compute.

Humanoid joints create a particular problem: cables routed through rotating connections are a common mechanical failure point. WIRE (Wireless Interface for Robotics) removes that constraint with short-range RF connectivity at 1 to 11 Gbps. No physical cables cross the joint.

At the limb level, sub-nanosecond PWM resolution produces smooth, human-like joint movement, while embedded security adds point-to-point link authentication within the robot body. Every internal communication path is verified as trusted.

Where Is NXP Deploying Robotics Silicon Today?

NXP and NVIDIA are collaborating on safety certification for robotic systems. NXP contributes real-time networking, motor control, and body-level safety; NVIDIA provides brain-level compute. The partnership includes full robot-body simulation within NVIDIA’s environment, supporting sim-to-real transfer for development and validation.

Customer deployments span all three form factors. Jetsy uses UWB-based precision landing for hospital medication delivery drones. A logistics drone customer adopted NXP’s RT1176-based flight control reference design. JD Logistics added SIL2 safety capability to its AMRs using MCX and AP-based devices. In China, a quadruped robotics company moved from proof of concept to production on motor control.

Boston Dynamics began working with NXP on AI noise cancellation, then expanded into facial recognition, vision-based following, and UWB tracking. Boston Dynamics robots are physically deployed in NXP’s France laboratory for ongoing development.

NXP has moved from selling individual components to offering reference implementations that span multiple layers of the stack. Silicon and software originally developed for software-defined vehicles are being adapted for robotic platforms, and customers with mature smart factory and Industry 4.0 programs are the most active in co-developing these solutions.