Photo/Boston Dynamics

This week, humanoid robots have once again become a hot topic.

Boston Dynamics, a veteran in the industry, has just retired its hydraulic version of the Atlas robot, only to revive it in a brand-new, electrified form.

According to Jim Fan, a senior scientist at NVIDIA responsible for the development of foundational models for humanoid robots, the astonishing design of the electric version of Atlas represents a breakthrough. He suggests that previous humanoid robot development was too obsessed with achieving “human-level” capabilities, which inadvertently set limits. Why not design robots that surpass human abilities from the start? The new Atlas has clearly broken free from the confines of merely imitating human movement.


On the same day that the electric Atlas was unveiled, the startup company Mentee Robotics also showcased the prototype of its first humanoid robot, Menteebot, claiming that AI has been integrated into all operational layers, making it a “guidable” personalized AI robot. The company explains that Menteebot’s perception is powered by pure visual sensing and driven by “unprecedentedly” flexible dedicated electric motors. With built-in AI algorithms and large language models providing the “brainpower,” it can learn and complete new tasks quickly through reasoning, not limited to a set of commands.

With the support of AI large models, the popularity of humanoid robots continues to rise. The IEEE Spectrum magazine, published by the Institute of Electrical and Electronics Engineers, has declared this year as the “first year of humanoid robots.”

Although we do not yet know the AI capabilities of the electric Atlas, Boston Dynamics has long recognized the importance of AI. Its founder, Marc Raibert, established the Boston Dynamics AI Research Institute in August 2022 to explore how robots can complete complex challenges with little or no human guidance.

Recently, the startup Figure AI has made a significant impact in the tech world with its excellent integration of humanoid robots and AI. In March, the company released a demonstration video showing its humanoid robot, Figure 01, understanding human verbal instructions and performing corresponding grasping and placing actions. The process is seamless and conversational, giving off a strong sci-fi vibe.

In February of this year, Figure AI, which was established less than two years ago, received $675 million in funding from a prestigious team composed of venture capital branches from Microsoft, NVIDIA, OpenAI, and Amazon founder Jeff Bezos. This funding raised the company’s valuation to $2.6 billion.

As the “arms dealer” of chips in this round of AI fever, NVIDIA established the GEAR Embodied Intelligence Lab in February to promote the development of AI technologies such as large models from the virtual world to the real world. At its annual GTC developers’ conference in March, the company introduced a humanoid robot project called GR00T, providing “general humanoid robot foundational models” to humanoid robot manufacturers, continuing to invest in the humanoid robot field.

Google DeepMind has previously released “RoboCat,” which has the ability to “self-improve,” and the VLA model “RT-2,” which integrates large language model capabilities. This has further accelerated the intelligentization of robots, sparking a new wave of innovation.

In addition to the aforementioned companies, other active players in the humanoid robot field include Tesla, Agility Robotics supported by Amazon, the Norwegian company 1X invested in by OpenAI, and Apptronik based in Texas.

Goldman Sachs released a “Humanoid Robot Special Report” in March, pointing out that the increasing demand for humanoid robots mainly comes from the need to replace high-risk occupations, especially in manufacturing (such as electric vehicle assembly and component sorting) and special operations (such as hazardous and dangerous tasks).

Moreover, technological innovation and cost reduction are also important drivers for humanoid robots.

Everbright Securities notes that the powerful algorithms of AI large models can greatly improve the efficiency of training robots. Additionally, humanoid robots, as a new type of intelligent terminal, are also an important downstream application of AI technology.

Goldman Sachs also mentioned in the report that the “end-to-end” learning capability and the empowerment of multimodal AI large models have significantly improved the technical feasibility of humanoid robots. Furthermore, the material cost of high-specification robots has dropped from about $250,000 in 2022 to $150,000 in 2023, a decrease of 40%.

The use case of Agility Robotics’ humanoid robot Digit is a strong testament.

In October 2023, Amazon tested using Digit in its Seattle warehouse to move and sort standardized plastic product boxes. Emily Vetterick, the engineering guide for Amazon’s global robotics business, believes that in older warehouses, various obstacles such as wire cover plates, anti-slip plates, cardboard, wooden frames, etc., can make wheeled robots difficult to maneuver, but walking robots can easily pass through.

According to the IEEE Spectrum magazine, the cost of Digit is about $250,000 per unit. If it can achieve the company’s target of 20,000 hours of working time, then its hourly cost is only $12.50, which is much lower than the $30 per hour comprehensive labor cost (not the take-home pay for employees) that Amazon spends on warehouse workers.

Digit is one of the humanoid robots closest to large-scale commercial adoption. Agility Robotics currently plans to build 10,000 robots per year for use in the logistics industry.

Most other humanoid robot companies are turning their attention to the automotive manufacturing industry.

Boston Dynamics, which has been acquired by Hyundai Motor, states that the main purpose of the electric Atlas is to create the next generation of automated production capabilities for Hyundai Motor. Figure AI’s first customer is BMW, which plans to explore the use of the Figure 01 robot on its assembly line in North Carolina, a plan that may come to fruition within the next 12 to 18 months. Texas-based Apptronik has also reached an agreement with Mercedes-Benz to jointly explore the application scenarios of humanoid robots on production lines.

Goldman Sachs estimates that under the base case scenario, the global market size for humanoid robots will reach $1.5 billion in 2025, $12 billion in 2030, and $38 billion in 2035, with shipment volumes of 20,000, 256,000, and 1,378,000 units, respectively.

Despite the promising outlook, humanoid robots still face many challenges before they can be widely applied.

The IEEE Spectrum believes that currently, the capabilities of sensors and computing in humanoid robots are sufficiently strong, but the actuators that convert energy into mechanical energy remain complex and expensive, and batteries are also unable to meet the power requirements for a full work shift. In addition, to apply humanoid robots on a large scale commercially, a robust supply chain system and service facilities need to be established, which cannot be achieved overnight.

Disclaimer: The content and data in this article are for reference only and do not constitute investment advice. Verify before using.

Editor: Alexander