Unitree verifies a new trend: the core battlefield of embodied intelligence is not just the models.

```

The competition in embodied intelligence is entering a new stage. With Unitree Technology launching the WVLA2.0 embodied large model and completing a live demonstration without remote control, the industry is becoming increasingly aware that the core barrier in this race is not simply dependent on model size, but rather a full-stack capability covering low-latency architectural design, software-hardware integration, and accumulation of native data.

According to a research report released by Nomura International on June 28, analysts conducted an on-site visit to Unitree on June 15. In the demonstration, the G1 robot equipped with WVLA2.0 (World-model Vision-Language-Action) autonomously completed six consecutive tasks in a disturbed conference room environment without remote control, with a reasoning closed loop of about 90ms, equivalent to ten iterations per second. This is the first version developed by Unitree over two years with potential for commercial deployment. Management listed industrial manufacturing—joint motor assembly, material loading and unloading, and fixture handling—as the earliest commercial application scenarios, and sees massive real-world operational data from a global fleet of robots as a core asset.

The Nomura report also outlined NXP’s NeuralAxis architecture framework released at COMPUTEX 2026. The framework was proposed under the leadership of NXP President and CEO Rafael Sotomayor, and its core concept highly aligns with Unitree’s engineering path—that the true bottleneck of physical AI lies not in the inference scale of language models, but in whether a human-like spinal reflex, with latency as low as 40ms, can be built at the edge control layer.

The direct implication for investors is: The competition in embodied intelligence is evolving from “whose model is stronger” to “whose system is more complete.” The moat built by Unitree through full-stack in-house integration and the advantage of real-world data is something pure cloud model vendors cannot easily replicate.

NeuralAxis: Redefining the System Architecture Boundaries of Physical AI

NXP’s NeuralAxis architecture takes inspiration from the human nervous system and breaks down physical AI control logic into three decoupled yet coordinated layers: a reasoning layer analogous to the cerebral cortex (latency around 300ms), a coordination layer similar to the cerebellum (responsible for motion control and balance), and a reflex layer analogous to the spinal cord—latency as low as 40ms, deployed near the actuators at the edge.

This framework has the most far-reaching implications for humanoid robots.

NeuralAxis proposes replacing a centralized “central brain” with distributed reflex processors—deploying local autonomous decision-making capabilities in joints, hands, and feet, achieving local execution of grip control, ankle balance, and other movements, and restoring balance, grasping, posture, and gait within 40ms. Decoupling reasoning and motion control also enables the robot to maintain movement stability while continuously adding new skills.

The commercial potential of this framework is equally noteworthy. Nomura’s industry research indicated that, compared with traditional automation solutions, the NeuralAxis framework can bring significant improvements in manufacturing efficiency, and sales of diagnostic robots are also expected to increase substantially. In addition, the same architecture can compress drone end-to-end latency to under 20ms, and can layer the control logic of software-defined vehicles for reasoning, coordination, and safety-critical regional execution.

WVLA2.0: Implementation Path for Model Fusion and Software-Hardware Collaboration

The technical route of Unitree’s WVLA2.0 demonstrates a clear divergence from mainstream industry approaches.

Most similar solutions bet on end-to-end pure VLA (Vision-Language-Action) generation, whereas WVLA2.0 integrates the world-model action (WMA) model’s predictive capabilities with VLA action generation, fully upgrading high-level task understanding, 2D/3D spatial semantic reasoning, dynamics-constrained action generation, and anti-interference capability.

In terms of perception, the system integrates four parallel visual streams: one RealSense depth camera, one Livox MID360 LiDAR, and two lateral cameras to construct 360-degree spatial representation, keeping position update latency under 10ms under interference. For software-hardware co-design, post-inference action parameters are communicated via CAN bus to the G1’s 23 degrees of freedom joints. With Unitree’s self-developed “cerebellum” motor control module, the single-arm grasping location error for objects under 2kg can be controlled within 5mm.

In computational architecture, WVLA2.0 compresses edge compute power to below 100 TOPS, running entirely on the G1 EDU equipped with NVIDIA (NVDA US, unrated) Jetson Orin NX, with no reliance on the cloud. Management says this design avoids task interruption risks resulting from network latency or disconnection.

Data Paradigm Shift: “No Native Data Collection” Going Mainstream

The shift in data collection modes is another key signal from this report.

Unitree’s demonstration showed that, without remote-control intervention, the G1 can autonomously complete multiple consecutive tasks in a disrupted environment in a single session, indicating that “no native data collection” is becoming the mainstream paradigm for generating embodied intelligence data—in other words, robots accumulate data through their own perception and decision-making, rather than relying on manual remote-tagged data.

Nomura’s industry research also pointed out current limitations: the system still has blind spots and rear perception gaps, slow execution speeds, insufficient fine manipulation accuracy, and lacks ongoing quantitative success rate benchmark data. These shortcomings also define the near-term boundaries for commercial deployment.

Management has thus formulated a phased rollout plan: industrial manufacturing (joint motor assembly, loading and unloading, fixture handling) is identified as the earliest landing point due to Unitree’s own factories providing a data closed loop; next comes logistics sorting and flexible 3C assembly; home and medical care scenarios are set as longer-term goals due to the significantly greater difficulty of open, unstructured environments.

Full-Stack Integration: Two Dimensions of Unitree’s Differentiated Barrier

The core conclusion of the Nomura report can be summarized as: In the commercialization process of embodied intelligence, model capability is certainly important, but it is not the sole decisive variable.

Unitree’s management defines the company’s differentiated competitiveness along two dimensions: First, the full-stack, in-house integration capability from perception through models to motion control; second, the accumulation of massive real-world operational data from a global robot fleet. These two assets reinforce each other—in-house hardware generates exclusive data, which feeds back into model iteration, creating a closed loop difficult for cloud model vendors to penetrate.

From a market competition perspective, the landing logics of the NeuralAxis framework and WVLA2.0 both point to the same conclusion: the core battlefield of embodied intelligence is unfolding simultaneously at the system architecture and data layers. For investors, evaluation dimensions for players on this track need to extend from purely “model capability” to a broader, more complete system integration capability and the scale of real-world data accumulation.

~~~~~~~~~~~~~~~~~~~~~~~~

The above highlight content is from Chasing Wind Trading Desk.

For more detailed explanations, including real-time interpretation and front-line research, please join [Chasing Wind Trading Desk・Annual Membership]

Risk Warning and DisclaimerThe market has risks, and investment needs to be cautious. This article does not constitute personal investment advice, nor does it take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views or conclusions in this article fit their specific situation. Investment is at your own risk. ```