Six Trends in the Development of Embodied AI and Humanoid Robots

In 2025, the global humanoid robot industry is at a critical turning point from technology verification to scenario penetration, and the potential market posed by various scenarios such as industry, service, special, and family exceeds tens of trillions of yuan. However, the humanoid robot market still faces three major bottlenecks in its scale-up: first, the cost system of humanoid robots has not yet been broken through; second, there is a generational gap in the intelligence level; and third, the supply of data elements is seriously insufficient. ResearchInChina has expounded the technical routes and product matrices of 21 leading Chinese and American companies and their signature products, analyzes the competitiveness of their humanoid robots, the cost reduction strategy of the next-generation products, and the direction of product evolution.

Trend 1: The humanoid robot market has undergone three technological iterations

The evolution of the humanoid robot industry essentially reflects the deepening of human cognition of "intelligent life forms". From early experiments with mechanical skeletons to today's autonomous decision-making capabilities based on AI foundation models, technological breakthroughs are gradually eliminating the boundaries between "machines" and "humans." So far, the development of the humanoid robot industry can be divided into three important stages, namely the initial exploration stage, the technology accumulation stage, and the stage of AI foundation models reconstructing cognitive decision-making systems.

Initial exploration stage (late 1960s-late 1990s): Dynamic walking theory to build a mechanical skeleton

During the initial exploration stage from late 1960s to late 1990s, the United States, the European Union, and South Korea focused on the kinematics and dynamics principles of bipedalism. The "Dynamically Stable Legged Locomotion" proposed by Professor Marc Raibert of the United States offered the basic technical outline. At this stage, Boston Dynamics was a typical veteran (1992).

Technology accumulation stage (early 2000s-2022): Sensors empower physical world interaction

During the technology accumulation stage from early 2000s to 2022, the industry focused on the deep integration and system integration of sensing and intelligent control technologies. At this stage, robots not only optimized basic motion control, but also made breakthroughs in the perception of basic information about the surrounding environment, and could adjust actions based on simple judgments. The technological accumulation laid a solid foundation for the rapid development of humanoid robots. For example, the bipedal Atlas robot demonstrated by Boston Dynamics in 2013 can walk, run, dance, carry and even perform difficult movements in complex terrains, marking a key step for humanoid robots to move towards more complex and intelligent application scenarios. Many players in China and the United States dabbled in the arena, including UBTECH (2012), Agility Robotics (2015), Unitree Robotics (2016), and Apptronik (2016).
 
New Embodied AI era (2022-present): AI foundation models reconstruct cognitive decision-making systems

After 2022, the humanoid robot industry witnessed a historic turning point - breakthroughs in AI foundation models such as OpenAI GPT-4 and Google RT-2 granted robots semantic understanding, task decomposition and autonomous decision-making for the first time, pushing the industry into a new era of Embodied AI. Through an end-to-end foundation model, Tesla Optimus can autonomously learn complex tasks (such as object classification and path planning) only from human demonstration data, improving the decision-making accuracy by 60% and significantly reducing the cost of manual programming. At the same time, the physical simulation engine built on the NVIDIA Omniverse platform supports the "virtual training - physical verification" closed loop by accurately simulating the dynamic characteristics of real scenarios, improving the robot development efficiency by 3 times and reducing trial and error costs by 80%.

Amid the technological innovation, start-ups such as Figure AI (2022), LimX Dynamics (2022), AgiBot (2023), and Galbot (2023) emerged, delving in vertical scenarios such as industrial manufacturing, logistics warehousing and home services. Traditional OEMs (Toyota, Hyundai, GAC, Chery, etc.), emerging OEMs (Tesla, Xpeng, Xiaomi, etc.), and AI infrastructure companies (Nvidia, DeepMind, Huawei, etc.) have accelerated their strategic layout and seized the market by virtue of their advantages in technology research and development, manufacturing processes or ecological resources. They work together to inject robust momentum into the booming humanoid robot industry.

Trend 2: Humanoid robots involve 4 types of players, with sports, scenarios, manufacturing and AI being key factors in the arena

Humanoid robot body design, manufacturing, and integration are the core links of the humanoid robot industry chain and the key to the industrialization and commercialization of humanoid robots. At present, the humanoid robot body industry is still in the exploratory stage. Humanoid robot body players can be roughly divided into four categories according to their original attributes: veteran robot companies, native robot companies, automotive OEMs, and start-ups. The four types of players are tackling industrialization difficulties in different ways - veteran companies break through the limits of sports, native companies consolidate the foundation of scenarios, automotive OEMs reshape the manufacturing paradigm, and start-ups lead AI integration - they jointly promote the humanoid robot industry to cross the "0-1" inflection point and open up a potential trillion-yuan "human-machine integration” market.
 
Trend 3: ROBOTERA STAR1 and Xpeng Iron are leaders in full body freedom  

With stronger sports flexibility, ROBOTERA STAR1 and Xpeng Iron are leaders in full body freedom. Unitree Robotics H1 has 19 DoF, Walker S1 has 41 DoF, Yuanzheng A2  40+ DoF, and Figure 02 only 16 DoF, showing significant differences. By rated joint torque, ROBOTERA Star1 boasts 400 N?m with a high power output; Unitree Robotics H1 360 N?m, CyberOne 300 N?m, and Galbot (G1) only 120 N?m.

Overall, ROBOTERA Star1 and Xpeng Iron have the most movement flexibility and load with top-notch degrees of freedom and joint torque, so that they are good at complex and high-load tasks. Yuanzheng A2 has a high degree of freedom (40+ DoF), outstanding movement flexibility, and balanced joint torque, suitable for medium-complexity tasks; Unitree Robotics H1 and CyberOne have obvious advantages in joint torque but relatively low degrees of freedom, and are more suitable for scenarios with large loads and simpler movements; Digit-4 and Figure 02 have low degrees of freedom and torque, and mainly fit for basic simple tasks; CL-1 and GoMate have medium degrees of freedom and joint torque, basic movement flexibility and task execution capabilities, ideal for routine operation scenarios.

Trend 4: Most humanoid robots have a range of about 2 hours, and a few reach 8-12 hours

Most humanoid robots have a range of about 2 hours, which is mainly limited by the insufficient battery energy density and the high energy consumption of joint drive. For example, Unitree Robotics H1 can work for an hour, UBTECH Walker S1 and Xpeng Iron 2 hours each. Some companies have achieved breakthroughs through structural optimization or battery technology innovation. Leju KUAVO-MY and Apptronik Apollo have a range of 4 hours after a single recharge; Agility Robotics Digit - 4 can last for 8 hours, and its next generation is expected to work for 12 hours; Galbot (G1) boasts a range of up to 12 hours with its wheeled dual arms and omnidirectional mobile chassis design, suitable for industrial scenarios. In the short term, energy consumption can be reduced through algorithm optimization and modular design. In the long term, high energy density technologies such as solid-state batteries and sodium-ion batteries can help break through bottlenecks.

Industrial scenarios have strict requirements for robot range (8-12 hours), and currently only some products are close to the standard; due to the fragmentation of tasks, home services require moderate robot range. For example, Xiaomi CyberOne can actually work for 3.5 hours, and GAC GoMate adopts all-solid-state batteries and variable wheel foot design to reduce energy consumption by 80% and increase range to 6 hours. In the future, the progress in range technology will lay the foundation for humanoid robots to cover a wider range of application scenarios.

Trend 5: Humanoid robots will evolve towards lightweight, multi-dimensional perception, and anthropomorphic motion.

Tesla Optimus features "human-like flexibility, industrial reliability, and AI autonomy" as a general humanoid robot, becoming the core terminal of Tesla's "hardware as a service" strategy. The evolution trend of Optimus is as follows:

(1) Lightweight design Magnesium alloy (density: 1.72g/cm3) and carbon fiber composite materials reduce the weight of Gen 2 has been reduced from 73kg to 63kg while ensuring structural strength, improving energy efficiency and movement flexibility, catering to the long-term operation and agile operation of the robot, and also laying the foundation for more scenario applications (such as home services).

(2) Multidimensional perception: For touch and force perception, fingertip pressure sensors, sole tactile matrix, 6-dimensional ankle force sensors and wrist multi-dimensional force sensors have been added to achieve more accurate contact force perception and balance control, adapting to complex scenarios. Force/torque: 6-dimensional ankle force sensors (dynamic balance control) + multi-dimensional wrist force sensors (real-time adjustment of operation force)

(3) Motion optimization: The walking speed increases from about 6 km/h to about 8 km/h (an increase of 30%), and the sense of balance and body control are significantly improved. The optimized actuator configuration (the number of rotational joints increases from 20 to 28, and the number of linear joints rises from 8 to 14) and motion algorithm make the robot more agile and stable, enabling it to perform complex movements such as squats and single-leg yoga, evolving towards a movement pattern closer to that of humans.

(4) Intelligence and algorithm advancement:
Computing power: Equipped with Dojo D1 (362 TOPS computing power), end-to-end training (video input → control output) is supported
Neural network: Preset action programming has evolved into AI autonomous decision-making, with joint control instructions directly generated through visual signals
Training method: Based on reinforcement learning of Tesla's factory data, the walking gait and operation strategy are dynamically optimized

(5) Actuator system upgrade: Quantity and complexity: The number of rotational joints and linear joints has increased, and the hand actuators have been upgraded from a simple grasping structure to 11-degree-of-freedom dexterous hands (3 degrees of freedom per finger + 2 degrees of freedom for the thumbs), improving movement flexibility, diversity and operation accuracy.

Trend 6: 2025 is the first year of mass production for structured scenarios, and home scenarios will be the focus in the next 5 years

2025 will be the first year of mass production for industrial manufacturing and automobile manufacturing

For the market demand side, humanoid robots can efficiently undertake high-precision and repetitive operations that are difficult for automated equipment to complete in industrial manufacturing, and promote full automation of industrial production. Structured scenarios such as industrial manufacturing, logistics and warehousing with strong standardization have low technical barriers, so model training is relatively easy. Based on this, most humanoid robot companies regard structured scenarios such as industrial manufacturing, automotive intelligent manufacturing, warehousing and logistics, and security inspection as the "first arena" for commercialization.

The penetration of humanoid robots follows the process from "structured scenarios → semi-structured scenarios → unstructured scenarios → general scenarios". Home scenarios will become the layout focus of representative humanoid robot companies in the 2025-2030

With a huge base of 1.6 billion households worldwide, the rigid demand for care and companionship incurred by aging, and average daily demand for more than 10 hours of housework, home scenarios constitute the main increment of the trillion-dollar consumer market. As the ultimate interactive entrance to the smart home ecosystem, the layout in home scenarios essentially embodies the strategic competition for the right to define the future "human-machine integration" lifestyle. Both the technological paths of American industrial robot giants and the ecological strategies of Chinese all-scenario players regard complex human-machine collaboration in home environments as the core arena. Although the stringent requirements of unstructured scenarios for robots' semantic understanding and dynamic decision-making mean that mass production will be the result of 5-10 years of technological iterations, this field has long become a strategic stronghold for future smart terminals.