Many people ask me, "How many more years will it take before we see humanoid robots in real life?" It's hard to give a definite answer to this question. According to the optimistic Elon Musk, mass production could be achieved within two years. However, looking back at the research process of humanoid robots over the years, due to limitations in application scenarios and extremely high requirements for composite technologies, it may take at least three to five years to explore a smooth business model.
Do we really need humanoid characters? How can we use them? Is there any need to develop them? These are highly controversial questions, and I'd like to share my views today.
In conclusion, the significance of humanoid robot research lies in: 1. Developing robot forms that are most adapted to human society, realizing the ultimate goal of machines replacing humans, but new breakthroughs are needed for specific scenarios; 2. Humanoid robots have extremely high requirements for integrated hardware and software technologies, and robot companies that focus on humanoid robot research and development can master the most comprehensive technology stack.
The first humanoid robot was created before the 21st century.
[I. Developing robot forms best suited to human society, but new breakthroughs are needed for application scenarios]
The goal of humanoid robots is to become the iPhone of the robotics world—the ultimate, universal form. Let's break it down: the required capabilities include movement (lower limbs - chassis/quadruped/bipedal), manipulation (upper limbs - single arm/double arm), and interaction (face). Through permutations and combinations, robot forms that meet these conditions can include:
- The simplest form, similar to the Google Everyday Robot, with a mobile chassis and a single arm;
- Best athletic ability, centaur, quadruped + one or two arms;
- The most scalable design, humanoid, with two legs and two arms, which is the direction Tesla and Boston Dynamics are heading.
From left to right: Google, Boston Dynamics, Tesla
If a robot needs to be developed as an everyday assistant, the simplest solution might be the Everyday Robot, capable of handling most tasks in production and daily life. However, this form has significant limitations when it comes to tasks requiring climbing stairs or using both arms. Since all aspects of human society are designed for human action, humanoid robots may be the most suitable form for human society, and once the technology matures, they can be seamlessly integrated into all environments.
In addition, let's look at some data: In 2021, China had a labor shortage of 11.8 million people, the United States had a labor shortage of 7.1 million people, and the global population is expected to peak in the middle of this century. From the perspective of filling the labor shortage, a type of robot is needed to replace workers. We have drawn a table to measure the order of replacement. In summary: given that China's labor costs are still relatively low, general-purpose robots will first replace skilled workers with certain learning and wage barriers, and then gradually penetrate into other quadrants.
First, it will replace skilled workers with certain learning and salary barriers, and then gradually penetrate into other quadrants.
So where can humanoid robots be used in the near future? Taking UBTECH as an example, the first batch of humanoid robots has been widely used in education and greeting. The company filed for listing on the Hong Kong Stock Exchange in 2023, but it is difficult to generate incremental market opportunities here.
Second-generation robots have the potential to gradually penetrate various aspects of production and daily life. For example, Tesla's humanoid robots will be used on its car production lines. It's important to emphasize that in the automotive manufacturing process, traditional robotic processes include stamping (machine tools), welding, palletizing, and painting, where industrial robots are already saturated. However, many processes still rely on manual labor, such as material handling between processes and the installation of car doors and seats. These are tasks that traditional industrial robots cannot perform, and these are areas where humanoid robots can further penetrate the market.
As the industry with the highest penetration of industrial robots, automobile manufacturing production lines still have a large number of manual processes.
Furthermore, there is no consensus yet on the cost and selling price of humanoid robots.
In 2017-2018, the internationally recognized cost of a humanoid robot was within 300,000 RMB, but it has now been reduced to a minimum of 100,000-150,000 RMB.
The difficulty in reducing costs lies in the all-purpose nature of the Tesla Optimus, which prevents it from focusing on a specific scenario. It's said that the current material cost for the Tesla Optimus is $200,000, with a target price of $20,000 after mass production by 2025-2027 (estimated cost of 70,000-100,000 RMB).
II. Humanoid robots require extremely high levels of integrated hardware and software technology; robotics companies that focus on developing humanoid robots possess the most comprehensive technology stack.
One limiting factor for the limited practical application of humanoid robots is their extremely high technological requirements, particularly the integration of motion control software and the robot's hardware. This results in most demo units being unable to stand, walk steadily, or move stiffly, making it difficult to complete tasks. Note: Boston Dynamics is discussed as a special case. While their products can perform parkour, they prioritize enhancing movement capabilities regardless of cost, resulting in extremely high power consumption and noise levels, making mass production impossible.
Humanoid robot hardware performance indicators
Let's take a look at some of the major technical challenges of humanoid robots:
Technical challenge 1: Balancing overall power consumption and battery life
In terms of battery life, most humanoid robots can only last for 1-2 hours at present, but it is predicted that this can be increased to 20 hours in the future. In contrast, human workers can work continuously for 7-8 hours, so robots have not yet reached a usable level.
The reason for this is that each joint of a humanoid robot requires a large motor, resulting in high overall power consumption and short battery life. This necessitates new structural and algorithmic designs.
Technical challenge 2: Motor ability and gait control
In particular, there is a shortage of talent in gait control (locomotion) in China. The following is a summary of the technical paths for centralized motion control, which are still under exploration and the landscape is yet to be determined.
Comparison of overall machine control technology paths
Technical challenge 3: Core components, integrated joints
Comparing the cost structure of new energy vehicles and humanoid robots: battery costs have decreased from 38% to 2%, while motor and electronic control costs have increased from 12% to 39%. The current pain points are high costs and high customization requirements. In the overall cost, the motor accounts for 40-50%, and the reducer accounts for 20%.
The cost of batteries for humanoid robots decreased from 38% to 2%, while the cost of motors and electronic controls increased from 12% to 39%.
Controllers and actuators can be domestically produced, but spare parts for special components need to be imported. A balance needs to be struck between quality, price, and delivery time (domestic customization takes 2-3 months, while imports take 8-10 months). Manufacturers already producing motors and reducers, having only experience in manufacturing machinery , may create new upstream opportunities in the future.
Bionic components need to be custom-made by research teams, including electronic skin, artificial muscles, and artificial cartilage.
Domestically sourced components that are easily subject to restrictions include new materials for mechanical parts, circuit board-related sensors and controllers, communication module chips , and circuit management chips.
Tesla's 6 Custom Motors and Drives
A Roundup of Domestic and International Humanoid Robot Companies
Companies that followed the humanoid technology path have arguably set themselves one of the most difficult problems in the robotics industry. Their progress has varied, and the following is a roundup of key companies:
- A benchmark for overseas mass production: Tesla bot
- Overseas benchmark for athletic performance: Boston Dynamics
- Starting with large models: 1x, Ameca
- Other overseas startups: figure, Agility Robotics, Sanctuary AI, prosper
- Domestic companies: UBTECH, Xiaomi CyberOne, Tencent Robotics X, Iron Man, university labs (Huang Qiang from Beijing Institute of Technology, Zhao Mingguo from Tsinghua University, Xiong Rong from Zhejiang University, Xing Tian from the University of Science and Technology of China), etc.
Comparison of parameters of some humanoid robot products
Tesla Bot, Optimus
- The concept was released in August 2021, and it was developed in 6 months. It weighs 73kg, has a 2.3kWh battery, an 8-hour battery life, and 28 degrees of freedom.
- The goal is to mass-produce a large number of reusable and replaceable parts, reduce costs, and use software to improve hardware (e.g., reduce the number of cameras from 8 to 3).
- We have developed 6 types of electric actuators that can pull a 500kg piano and have a response speed of up to 10kHz, but they need to be paired with a harmonic reducer (joint, which uses rotational speed to exchange for torque).
- High torque density results in strong explosive power and low copper loss, so when the optimus output is large, the battery life is long and the heat generation is low.
In Tesla's latest video , the robot has already achieved: 1. End-to-end motion generation, teleoperation, and teaching functions; 2. High lower limb torque control frequency (estimated to reach 1000 Hz), making its steps light enough not to break eggs.
Boston Dynamics
- It began iterations in 1992, weighing 150kg, with a 3.7kWh battery and a 1-hour battery life. The goal was to demonstrate athletic capabilities beyond human limits.
- Drive: Hydraulic, a liquid compression pump generates high-pressure liquid, which performs work on the output pump. Advantages: high force, only one actuator required; disadvantages: complicated maintenance, slow response speed. Metal 3D printing manufacturing.
- Development history: In 1992, it separated from the MIT laboratory; in 2013, Google acquired it for $3 billion; in 2017, SoftBank bought it from Google for an undisclosed price; in the second half of 2020, the Spot robot dog was released for $74,500; in December 2020, Hyundai Motor acquired it for $921 million, and the company's annual expenditure was $150 million.
Norway/USA 1x
- Background: Founded in 2015, this Norwegian company initially developed service robots for the elderly and recently announced its entry into the security market.
- Team: CEO, Bernt Børnich , Norwegian. AI Lead, Eric Jang, Chinese-American, Senior Research Scientist at Google X for 6 years.
- Product: Priced at US$150,000, height 175cm, weight 76kg, single-arm load capacity 8kg, speed 22km/h.
Latest funding: After launching the development of a humanoid robot prototype called NEO, the company received investment from OpenAI and Tiger Global Management.
- OpenAI COO Brad Lightcap commented: The 1x investment is aimed at the maturity of robotic hardware technology and the expansion of a broader labor market with the support of AI.
British Engineered Arts, Ameca
- Starting by creating animated characters for the entertainment industry, they use light field cameras to scan real people and 3D print facial muscles.
- After Ameca integrates with ChatGPT, it can display fluent emotions and responses to human language.
figure.ai (USA)
- Timeline: Founded in 2022, the company released a promotional video for Robot No. 01 in March 2023, with plans to launch a test product within 6 months.
- Specifications: Similar to Tesla , 1.67m tall, 59kg, electric motor driven, 5h range.
- Application scenarios: warehousing and logistics. There are already 5 prototype machines, and 50 robots will be manufactured in 2024.
- Founder: Brett Adcock , serial entrepreneur, with a personal investment of $10 million. He recruited 40 technical experts from Boston Dynamics, Tesla, Google X, and the robotics lab IHMC.
Agility Robotics (USA)
Founded in 2015, with a CEO who is a PhD from CMU, it was the first to launch a commercial humanoid robot and is preparing to make its mark in warehouses.
- Raised $150 million in Series B+ funding round, with investors including Amazon , Sony, and Tokyo Electric TDK .
Canadian Sanctuary AI
Founded in 2018, it has a humanoid upper body and wheels on its lower body, and began testing in stores. It raised $60 million in Series A funding.
Prosper UK
- In 2023, a general-purpose robot demo was released, which can make beds (a challenging task involving the recognition and manipulation of flexible fabrics) and operate washing machines. The founder previously worked at OpenAI and Scale AI.
UBTECH
- According to the 2023 annual report of the Hong Kong Stock Exchange, the valuation is RMB 30 billion (PS 38x in 2021, 43x in 2022).
- From 2020 to Q3 2022, revenue was 740 million yuan, 820 million yuan and 530 million yuan respectively, with net losses of 710 million yuan, 920 million yuan and 780 million yuan respectively, and a cumulative sales volume of 500,000 units.
- Revenue breakdown: Education 68%, Logistics 10%, Consumption 16%, Other 6%
- Profit: Gross profit margin 31%, sales expense ratio 46%, R&D expense ratio 61%.
In summary, humanoid robots are currently limited by application scenarios, technology, and cost, and have not yet achieved widespread deployment. However, we believe that humanoid robots represent the ultimate form of robot capable of penetrating human society, enabling machines to replace or supplement humans. The process of industrialization will also spur the emergence of numerous upstream and downstream companies.
