Unitree's Wang Xingxing Explains Why They Built the 3.9-Million-Yuan Manned Transforming Mech: Large Robots Are an Unstoppable Industry Trend

[Core Event and Technical Overview]

In May 2024, Unitree Robotics officially launched the GD01 manned transforming mech with a starting price of 3.9 million yuan, marking the formal entry of a domestic embodied intelligence company into the large manned robot sector. At the 5th Global Digital Trade Expo held in Hangzhou from September 23 to 27, the GD01 appeared in the theme pavilion exhibition area, attracting numerous on-site inquiries but with no first unit delivered yet. The product adopts a hybrid bipedal and wheeled transforming configuration, equipped with multi-degree-of-freedom servo joint modules and torque sensor arrays, enabling dynamic balance control and posture switching in manned mode. Unitree positions it as a technology platform for high-end exhibitions, scientific research validation, and specialized scenarios, rather than a mass consumer product.

At the Digital Trade and AI Dialogue Forum of the expo, Unitree founder Wang Xingxing delivered a keynote speech titled "From Mechanics to Intelligence—An Evolutionary Theory of Embodied Futures," systematically articulating the strategic logic behind developing the manned mech. Wang Xingxing pointed out that as AI large models and underlying robotic hardware technologies gradually mature, the development and deployment of large robots and small robots are not in conflict but rather serve different scenario dimensions for technology validation. Large robots are an unstoppable industry trend—even if Unitree does not build them, other companies will enter this space. The core rationale for this judgment is that large platforms can carry higher power-density drive systems and more complex perception-computation units, providing an irreplaceable physical testing ground for validating embodied intelligence algorithms under extreme load conditions.

[Technical Principles and Core Breakthroughs]

The technical architecture of the GD01 manned mech integrates multiple cutting-edge engineering practices from the embodied intelligence field. At the underlying drive level, the core challenge for large manned platforms is the design and thermal management of high-power-density servo joints. Unlike traditional quadruped robots with single-joint peak torque in the tens of Nm range, the hip and knee actuators of the manned mech need to continuously output hundreds or even over a thousand Nm of torque while maintaining precise torque response bandwidth. This requires coordinated optimization of motor winding design, gear ratio configuration, and current-loop control frequency. Unitree had previously validated a joint module solution with 360 Nm peak torque on the H1 humanoid robot; the GD01 further upgrades the continuous power rating and heat dissipation efficiency on this basis, adopting a hybrid liquid-cooling and air-cooling thermal management architecture to ensure controllable temperature rise during prolonged manned operation.

At the perception and decision-making level, the GD01's embodied intelligence technology stack relies on an end-to-end architecture based on Vision-Language-Action (VLA) models. The mech is equipped with a multi-camera depth camera, LiDAR, and IMU combined perception system, achieving real-time 3D reconstruction of complex environments and obstacle semantic segmentation through multimodal feature fusion. In terms of control strategy, balance control employs a hybrid scheme of Model Predictive Control (MPC) and Reinforcement Learning (RL): MPC handles dynamic balance constraints and ZMP stability determination under manned conditions, while the RL policy network acquires adaptive gait generation capabilities on unstructured terrain through large-scale simulation pre-training. This hybrid architecture has accumulated over 100,000 hours of training in simulation environments, covering terrain conditions including slopes, stairs, and gravel roads. The manned mode introduces an additional safety redundancy layer, including torque limiting, emergency stop response, and operator intent parsing modules, ensuring physical safety boundaries during human-machine collaboration.

From a system architecture perspective, the GD01 represents a critical node in the evolution of embodied intelligence from pure algorithm validation toward strong physical-world interaction. The core technical barriers of a large manned platform lie not only in individual joint power or perception accuracy, but more importantly in system-level integration capability of the entire machine—including full-chain coordination of power management, thermal design, structural strength, control latency, and fault diagnostics. Unitree's choice to launch the manned mech at this point in time reflects that its internal confidence threshold for joint module reliability and control algorithm maturity has reached a level capable of meeting manned safety requirements. Meanwhile, the transforming configuration design means the system must smoothly switch between two dynamic models—bipedal walking and wheeled cruising—placing extremely high demands on state machine design and the robustness of control switching logic.

[Industry Background and Competitive Landscape]

From the perspective of global competition, the large manned robot sector is still in a very early exploratory stage with few participants. Tesla's Optimus Gen-2 focuses on a 1.73-meter-tall, approximately 73 kg non-manned general-purpose humanoid, targeting labor replacement in industrial manufacturing scenarios; Figure 02 is similarly positioned as a general-purpose bipedal robot, collaborating with BMW and other manufacturers for production line task validation; Boston Dynamics' electric Atlas continues the technical route of high-dynamic locomotion capability but has not ventured into manned forms. Unitree's GD01 choice of the manned transforming mech as a differentiated positioning essentially avoids direct competition with overseas giants in general industrial scenarios, instead building brand momentum in the high-end exhibition and research platform market. The 3.9-million-yuan pricing also clearly indicates this is not a product aimed at scaled deployment, but rather a strategic positioning closer to technology validation and brand benchmarking.

In the domestic embodied intelligence sector, Unitree Robotics' core competitive advantage lies in the joint module mass production capability and motion control algorithm experience accumulated during the quadruped robot era. From the A1, Go1, to the B2 series, Unitree has achieved scaled production and cost control of servo joints, with single joint module costs at an industry-leading level. This supply chain advantage provides the engineering foundation for rapid iteration of the H1 humanoid robot and the GD01 manned mech. Meanwhile, emerging companies such as Agibot Robotics, Galbot, and Robot Era are also accelerating their deployment of general-purpose humanoid robots, but most focus on small-to-medium load scenarios. Unitree's choice of a large manned platform is on one hand a natural extension of its technical capabilities, and on the other hand a search for a differentiated moat in the increasingly crowded general humanoid market. From an industry-wide perspective, by the second half of 2024, domestic embodied intelligence companies exceeded 50, but fewer than 10 truly possess full-stack in-house R&D capabilities from joint hardware to control algorithms, with Unitree among the minority with a complete technology loop.

Placing the manned mech within the broader context of the robotics industry's evolution, the emergence of this category is not an isolated event. Japan's Suidobashi Heavy Industry previously launched the Kuratas manned mech priced at approximately 1 million USD, and the U.S. company MegaBots built large combat robots that attracted widespread attention, but these projects ultimately failed to achieve sustained commercialization. The difference with Unitree's GD01 lies in its birth against the backdrop of an era where AI large models empower embodied intelligence—the mech is no longer a purely mechanical cockpit, but an intelligent agent carrier with environmental perception, autonomous decision-making, and human-machine collaboration capabilities. This paradigm shift in technology makes it possible for manned mechs to transition from sci-fi concept validation to intelligent platforms with practical functional value, but whether their commercialization can break through the predicaments of predecessors still depends on the dual validation of scenario closure and safety compliance.

[Developer and Industry Implementation Insights]

From the perspective of engineering implementation and commercialization, the primary challenge facing the GD01 manned mech is not technical realization, but precise positioning of target customer groups and closed-loop validation of use scenarios. The starting price of 3.9 million yuan restricts potential customers to extremely narrow niche markets such as high-end commercial exhibitions, theme parks, research institutions, and specialized industry applications. The fact that there have been numerous on-site inquiries but no first unit sold yet reflects that the market is still in the awareness-building stage for this category. For potential buyers, core concerns include the lack of safety certification standards for manned operation, the technical threshold for maintenance and the incompleteness of spare parts supply chains, and whether the functional value the mech can provide in actual use scenarios matches its high procurement cost. Currently, China has no dedicated safety regulations or operational qualification certification system for manned robots, meaning each GD01 delivery requires Unitree to provide deeply customized deployment support and operator training, significantly increasing the engineering cost per unit delivered.

For developers and research institutions, the GD01's potential value as a large-scale embodied intelligence research platform is worth noting. The mech's high payload capacity and rich sensor configuration provide a unique experimental environment for validating VLA models under large-mass-body dynamic conditions. However, Unitree has not yet released standardized SDK interfaces and simulation environment development packages for the GD01, similar to Boston Dynamics' early closed strategy with Atlas. If Unitree could open up simulation training environments based on Isaac Gym or MuJoCo along with real-machine deployment APIs, it would significantly lower the barrier for research teams and accelerate the formation of an algorithm iteration ecosystem. In terms of migration costs, the main obstacles for research teams migrating from small-to-medium humanoid platforms to the GD01 are the recalibration of dynamic parameters and reconstruction of safety constraint strategies, with an estimated adaptation period of 2 to 3 months. For commercial deployment, the most viable path in the short term is to partner with leading cultural tourism groups or science and technology museums, replacing direct sales with an operating revenue-sharing model to lower customers' initial investment threshold.

From a hardware memory and computing resource perspective, the GD01's onboard computing platform needs to simultaneously process real-time inference from multiple vision sensors, LiDAR point cloud processing, end-to-end policy output from VLA models, and state machine loops for safety monitoring modules. This places stringent requirements on the computing power configuration of the edge computing platform, estimated to require at least two high-power-consumption GPUs or dedicated AI accelerator chips, with the total computing platform power budget exceeding 500 watts. For teams hoping to perform secondary development based on the GD01, additional computing resource redundancy must be reserved to accommodate inference loads from custom models. If Unitree can provide modular computing expansion interfaces and standardized ROS2 communication middleware adaptation layers, it would significantly reduce the engineering barrier for third-party algorithm integration—this is also one of the key factors determining whether the GD01 can truly become an open research platform.

[Comprehensive Review and Key Points]

The launch of Unitree's GD01 manned mech is essentially a landmark event marking the embodied intelligence industry's transition from a locomotion capability race to a scenario-defined product stage. Wang Xingxing's judgment that large robots are an unstoppable trend is technically sound—the power density, perception bandwidth, and computational redundancy carried by large platforms provide exploration space at the physical limits for validating more complex embodied intelligence algorithms. But from a commercialization reality perspective, the manned mech at its current stage is closer to a role of brand momentum building and technology limit validation, rather than a sustainable scaled commercial category. Over the next 1 to 2 years, key observation metrics for this category include the timing of the first actual delivery, functional closure validation results in operational scenarios, and whether a second manufacturer will follow suit with a similar product. If only Unitree maintains this category, it will remain more of a technical benchmark than the starting point of an industry trend.

From a broader industry evolution perspective, the embodied intelligence industry will enter a phase of hardware convergence and algorithm explosion from 2024 to 2025. At the hardware level, the trend toward standardization and cost reduction of servo joint modules is already established, with single-joint costs expected to drop another 30% to 40% within 12 months, creating conditions for scaled deployment of humanoid robots. At the algorithm level, VLA models will gradually evolve from the current paradigm of simulation pre-training plus real-machine fine-tuning toward online learning and continuous optimization based on real-world data. The unique value of manned mechs as large platforms in this evolution lies in their ability to collect real dynamic data of large-mass bodies under complex contact interactions—data that has irreplaceable supplementary significance for training embodied intelligence policy networks with stronger generalization capabilities. Even if the GD01 itself struggles to achieve scaled sales in the short term, the data assets and engineering experience accumulated during its operation will continuously feed back into Unitree's core competitiveness across its general-purpose humanoid robot product lines.

Overall, the strategic significance of Unitree launching the manned mech outweighs its direct commercial returns. It sends three key signals to the industry: first, the hardware engineering capabilities of domestic embodied intelligence companies have reached the threshold capable of meeting manned-level safety requirements; second, technology validation of embodied intelligence is no longer confined to laboratory environments but is beginning to extend into extreme real-world scenarios; third, the competitive dimensions among robotics companies are expanding from a single product performance comparison to a comprehensive contest of technology platform coverage and brand narrative capability. For industry practitioners and investors, it is necessary to rationally distinguish between technology.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.