Nomura focuses on the commercialization of Physical AI in Chinese robotics: from the NeuralAxis architecture to Unitree WVLA2.0

Overview

Based on a visit to Unitree on June 15, 2026, this report discusses the commercialization path for Chinese robotics and Physical AI. Nomura divides Physical AI commercialization into two layers: first, the NeuralAxis system architecture and safety principles proposed by NXP, emphasizing layered control analogous to the human nervous system; and second, how Unitree WVLA2.0 converts this type of blueprint into a deployable product through world models, VLA, and hardware-software co-design.

Core views

The core view is that the key constraint on Physical AI is not simply expanding language-model inference capabilities, but building a low-latency, distributed, and more energy-efficient system for reflexive and coordinated control. NeuralAxis divides the system into inference, coordination, and reflex layers. The reflex layer is pushed down to execution points such as joints, hands, and feet, helping complete local autonomous decision-making and posture recovery within approximately 40ms. Unitree WVLA2.0 demonstrates a productization route by combining WMA world-model action capabilities with VLA end-to-end action generation, and implementing closed-loop execution through the G1 robot's hardware, sensors, and local edge computing.

Analysis framework

The report combines industry research with a company visit. It first explains the underlying logic of NeuralAxis as a system blueprint for Physical AI, then uses Unitree WVLA2.0 as a case study to analyze how commercial products based on this concept are deployed. The analysis focuses on system architecture, latency, sensor fusion, motion control, edge computing, data collection paradigms, and priority commercial scenarios.

Methodology notes

• System architectureNeuralAxis

  Three-layer neural-axis architecture

  This framework maps Physical AI onto three layers analogous to the human nervous system: an inference layer for high-level decisions, a coordination layer for motion control and balance, and a reflex layer close to the actuators that performs safety-critical actions with low latency.

• Model architectureWVLA2.0

  Integration of world models and VLA

  Unitree WVLA2.0 combines WMA's predictive capabilities with VLA's end-to-end action-generation capabilities to improve task understanding, spatial semantic reasoning, dynamics-constrained action generation, and disturbance resistance.

• Commercialization pathIndustrial-first deployment

  From controlled industrial scenarios to open environments

  Management believes robots will first be deployed in industrial manufacturing, such as joint-motor assembly, loading and unloading, and fixture handling, before expanding to logistics sorting, flexible 3C assembly, homes, and healthcare. The latter applications will come later because open and unstructured environments are more challenging.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Unitree

  Core case company, unlisted

  Strengths

  Full-stack in-house integration capabilities, the G1 robot hardware platform, WVLA2.0 model-integration capabilities, and large-scale real-world machine data from its global fleet.

  Weaknesses

  The report points to continuing issues including blind spots in rearward perception, relatively high noise, slow execution, insufficient precision in fine manipulation, and a lack of quantitative benchmarks for continuous success rates.

  Comparison

  Compared with companies focused solely on VLA or cloud models, Unitree's approach emphasizes coordination among world models, VLA, robot hardware, and motion control.

  Risks

  Moving from demonstrations to large-scale commercial deployment still requires validation of sustained success rates, reliability, cost, scenario generalization, and adaptability to complex open environments.

• NVIDIA CORP / NVDA.US

  Company related to the edge AI computing supply chain

  Strengths

  The G1 EDU uses the NVIDIA Jetson Orin NX to run WVLA2.0 locally, demonstrating the role of edge AI chips in low-latency robotics applications.

  Weaknesses

  The report provides no quantitative assessment of NVIDIA's rating, valuation, or earnings impact.

  Comparison

  Compared with cloud inference, local edge inference can reduce latency and lessen task interruptions caused by disconnections.

  Risks

  Whether demand for on-device computing in robotics can translate into meaningful revenue depends on production scale, cost constraints, and solution substitution.

• NXP / NXPI US

  Originator of NeuralAxis, not rated

  Strengths

  NeuralAxis emphasizes low-latency reflex processing, distributed control, and energy efficiency, aligning with the safety-critical control requirements of Physical AI.

  Weaknesses

  The report mainly discusses the architectural concept and provides no estimates of NXP's commercial revenue or investment rating.

  Comparison

  This approach shifts the focus of Physical AI from a centralized brain toward distributed reflexive control close to the actuators.

  Risks

  The implementation of the architectural concept still depends on the adoption pace of end customers in robotics, drones, and software-defined vehicles.

Key data

• Unitree visit date2026-06-15The report states that Nomura visited Unitree on this date to update its understanding of the company's latest progress.
• NeuralAxis reflex-layer latencyas low as 40msReflex processing is moved close to the actuators for local autonomous control.
• Reference latency for inference layer~300msNeuralAxis compares the inference layer to the cerebral cortex.
• Drone glass-to-glass latency targetwithin 20msThe report states that the same architecture can compress end-to-end visual latency for drones to within 20ms.
• WVLA2.0 sensor configurationRealSense depth camera, Livox MID360 LiDAR, two side camerasFour visual/perception streams are fused into a 360-degree representation.
• Position update time under disturbancewithin 10msThe report states that WVLA2.0 can complete position updates within 10ms under disturbance conditions.
• G1 joint degrees of freedom23 degrees of freedomMotion parameters are sent via the CAN bus to the G1's 23 degrees-of-freedom joints.
• Single-arm lifting weightsub-2kgThe report states that positioning error is controlled within 5mm when a single arm grasps objects weighing less than 2kg.
• Single-arm grasping positioning errorwithin 5mmThis demonstrates Unitree's motion-control and hardware-software co-design capabilities.
• Edge computing capacity ceilingbelow 100 TOPSAfter lightweight optimization, the system can run locally on the NVIDIA Jetson Orin NX in the G1 EDU.
• Closed-loop inference cycle~90ms per inferenceThis represents approximately 10 iterations per second, covering the perception, prediction, decision, and action loop.
• Nomura rating distributionBuy 57%, Neutral 41%, Reduce 2%Disclosure of the global equity research rating distribution of Nomura Group as of March 31, 2026.

Impact & implications

For investment research, the report reinforces two key themes in the Physical AI value chain: first, robot bodies and control systems require low-latency edge computing and safety-control architectures closer to the execution layer; second, companies with integrated capabilities spanning complete machines, motion control, sensor fusion, and closed-loop real-world machine data may have greater commercialization advantages than pure cloud-model providers. Edge AI computing platforms such as the NVIDIA Jetson Orin NX, NXP-related embedded-control concepts, and the progress of Chinese robot manufacturers in industrial deployment may all be key areas to monitor.

Risks

• Unitree WVLA2.0 still has perception blind spots and gaps in rear-area perception.
• Demonstrations continue to show relatively high noise, slow execution, and insufficient precision in fine manipulation.
• The report provides no quantitative benchmark for continuous success rates, and reliability for commercial deployment requires further validation.
• Open and unstructured scenarios such as homes and healthcare are significantly more difficult than industrial manufacturing.
• The commercialization of Physical AI may be constrained by hardware costs, edge-computing power consumption, supply chains, regulation, and safety liability.

What to watch

• Whether Unitree WVLA2.0 subsequently discloses continuous success rates, failure rates, and real factory deployment data.
• Progress in commercial orders for tasks such as joint-motor assembly, loading and unloading, and fixture handling in industrial manufacturing.
• Differences in deployment pace across logistics sorting, flexible 3C assembly, home, and healthcare scenarios.
• The penetration rate of the NVIDIA Jetson Orin NX and similar edge AI platforms in on-device robotics deployments.
• Whether NeuralAxis-style distributed reflex control is adopted by more robotics, drone, and software-defined vehicle solutions.