Humanoid robots still need to pass their 'internship'.

Imagine this scene: a workshop at dawn, rows of neatly lined humanoid robots whirring away on assembly lines. A sight that was once confined to science fiction movies just a few years ago, but is now a reality.

Recently, Figure, a humanoid robot enterprise, unveiled its second-generation Figure02, marking breakthroughs in both hardware and software. Equipped with a custom speech reasoning model from OpenAI, its investor, Figure02 enables natural language interaction. It has already entered BMW factories, capable of working non-stop for 20 hours.

Image source: Weibo screenshot

The announcement of this 'groundbreaking' product has once again propelled humanoid robots into the spotlight. From UBTech's (9880.HK) IPO at the end of last year, to Tesla's Optimus generation at this year's WAIC, to the recent launch of Figure02, humanoid robot development has progressed rapidly over the past two years.

But beneath the veneer of 'innovative technology,' what are the challenges in mass-producing humanoid robots? What obstacles must be overcome from R&D to commercial application?

Humanoid robots are back in the spotlight.

Figure, a humanoid robot enterprise backed by Microsoft, NVIDIA, and OpenAI, recently unveiled its latest model, Figure02, touted as the 'strongest on the planet.' According to 'QuJieShangYe,' compared to its predecessor, Figure02 underwent a redesign of its exterior, with the robot's skin now bearing the stress, adopting an exoskeleton structure that integrates power and computing wiring within the body, enhancing performance. The revamped Figure02's hands have 16 degrees of freedom, capable of withstanding forces comparable to those of humans.

On the software side, the new robot is equipped with a custom speech-to-speech reasoning model from OpenAI, allowing for direct voice interaction with humans.

Image source: Weibo screenshot

Robots have already infiltrated our daily lives, from large cleaning robots in malls to logistics robots on assembly lines. With specialized robots so advanced, why do we still need humanoid robots?

And why must they be 'humanoid'?

The answer is simple: only humanoid robots have the potential to be truly 'versatile.'

The world we live in is designed around human forms and interactions. While specialized robots excel in specific tasks (e.g., lifting heavy objects, screwing bolts, or sweeping floors), they become limited outside their designated domains.

Image source: CanStockPhoto

Humanoid robots, however, are not bound by scenarios, integrating seamlessly into all aspects of human life. Their high degree of freedom, autonomy, and versatility make them ideal for tasks like replacing humans in repetitive, complex, and dangerous industrial tasks, reducing accidents. We can consider current robots as preliminary forms leading to the ultimate form of humanoid robots.

To achieve this ultimate form, humanoid robots must excel in three areas: perception, movement, and cognition.

Take obstacle avoidance as an example. The robot must first 'perceive' the obstacle, then 'cognize' the best course of action (walk, jump, or step over), and finally 'move' to execute it. While seemingly intuitive for humans, this is a complex training process for humanoid robots.

Most current humanoid robots struggle with the 'cognition' aspect. While perception and movement can be improved through sensor layout and joint design, cognition remains incredibly complex.

Traditional training methods involve 'tagging,' where AI learns by associating labels with objects (e.g., various types of watermelons). This works well in limited domains but falls short in broader contexts.

Image source: CanStockPhoto

Large models offer a new training paradigm. Combined with sensor data, they enable robots to better understand the world's patterns.

For instance, retrieving a drink from the fridge requires different approaches for bottled cola and bagged milk. Large models teach robots the nuances, like gently lifting bagged milk compared to grabbing bottled cola.

Crucially, large models enhance robots' understanding of natural language. If you say, 'I'm hungry,' a trained robot might navigate to a food delivery app or even cook a meal. This deep understanding simplifies human-robot interactions.

In short, large models propel humanoid robot development to new heights.

With their support, 2023 was dubbed the 'Year of Humanoid Robots.'

Financing reflects this growth. According to 'ITjuzi,' there were 23 financing events in the humanoid robot sector in 2023, totaling 5.471 billion yuan, the highest in a decade. As of now, 2024 has seen 14 financing events, raising nearly 2.4 billion yuan.

Image source: ITjuzi

Product-wise, 2023 has witnessed a surge in both startups and enterprises. For instance, Xingdong Jiyuan, founded in 2023 and incubated by Tsinghua University's Institute for Interdisciplinary Information Sciences, focuses on embodied AI and humanoid robot technology. Kepler, also founded in 2023, combines self-developed algorithms with precise control for humanoid robots.

Image source: CanStockPhoto

Veteran robot maker UBTech, listed just before the new year, reignited industry capitalization as the 'first humanoid robot stock.'

UBTech and similar third-party suppliers drive the sector, often targeting specific use cases (e.g., industrial manufacturing for UBTech, household services for CloudMinds). As midstream players, they prioritize downstream orders and early market entry. For example, UBTech's Walker S has been tested in NIO's Hefei factory, and the company is in talks with leading domestic EV makers.

Image source: Weibo screenshot

Some downstream enterprises, including Tesla, are developing their humanoid robots.

Tesla's Optimus, a cornerstone of Elon Musk's 'Master Plan,' has evolved to perform basic yoga poses with coordinated arms and legs, maintain stability, and sort blocks by color using vision alone. It adapts to disturbances and corrects mistakes.

Image source: Weibo screenshot

Xiaomi unveiled 'Cyber One' in 2022 and recently invested in 'Xiaoyu Zhimao,' a embodied AI tech company, marking its first foray into the field. Honda, with over 30 years in robotics, showcased ASIMO shaking hands with the NYSE Chairman in 2002.

Unlike midstream suppliers, downstream enterprises prioritize practicality and branding, especially in the auto industry. Smart robot research sharpens their tech edge and captures cutting-edge trends.

At Xpeng's 1024 Tech Day last year, CEO He Xiaopeng unveiled PX5, a bipedal humanoid robot developed in-house over five months.

On August 12, He shared a video on Weibo of a robotic arm ordering a MONA M03, teasing AI and hardware advances soon to come.

Image source: Weibo screenshot

Practically, midstream players emphasize 'open sourcing' for revenue, while downstream focus on cost control. Musk plans to integrate Optimus into Tesla's production lines, replacing human workers with 24/7, fatigue-free 'perfect workers,' cutting costs and boosting efficiency.

Stable downstream humanoid robots can also commercialize, contributing to corporate revenue. Despite shared urgency to 'enter factories,' midstream and downstream players have slightly different priorities.

Both midstream and downstream players face challenges amidst the buzz.

Tesla's humanoid robot batch production plans have slipped. After promising Optimus mass production by 2024, Musk tweeted limited production next year, with full-scale production and client deliveries targeted for 2026—a nearly one-year delay.

Xiaomi's 'Cyber One' was last showcased at the World Robot Conference 2023, featuring the 2022 version with no major updates since.

'UBTech expects small-scale humanoid robot deliveries to auto factories by the end of 2024,' said Chairman and CEO Zhou Jian. However, the listed company faces a sliding market value, down over 70% from its March peak of 328 HKD per share to 86.25 HKD as of August 14, wiping out nearly HKD 100 billion in market cap.

Image source: Baidu Gushitong screenshot

So, what hinders humanoid robot development?

First, technical hurdles. Media reports cite four challenges identified by leading R&D experts:

1. Difficult hardware integration in hands and legs, requiring high peak power and drive capability.

2. Unstable walking algorithms; true human-like gait algorithms are elusive in academia.

3. Complex hand perception involving 21-26 degrees of freedom.

4. Sophisticated algorithms for arm coordination, hybrid intelligence control, and obstacle avoidance, requiring high-dimensional planning.

Image source: CanStockPhoto

Second, cost control. Mass production necessitates balancing hardware and software costs, optimizing economies of scale.

Before cost control matures, market pricing competition intensifies. In May, Unitree's G1 humanoid robot priced at just 99,000 yuan shocked the industry, where prices often exceeded one million yuan. Shortly after, Zhongqing Robotics released the SA01 professional bipedal robot at 36,500 yuan, targeting the research and education market.

Picture source: screenshot from Weibo

This polarization poses a dilemma for many robot manufacturers, which reflects the unclear demand in the downstream market for humanoid robots. This is also a problem with the commercialization of most high-tech products. Since high-tech products have extremely high requirements for technological innovation, the resulting products that meet the requirements are too expensive, so midstream manufacturers can only launch one cheap version after another.

In addition, the application of humanoid robots in unstructured scenes is limited; due to the lack of versatility, robots cannot flexibly respond to emergencies. The most urgent need for the commercialization of humanoid robots is a pressing and precise scenario; given the current situation, industrial scenarios under high-end manufacturing demands and daily life scenarios under the trend of population aging are the scenarios with the highest commercial value. However, there is still a lack of a product paradigm in the market to allow people to more intuitively see the commercial value of humanoid robots.

Humanoid robots still have a long way to go before they can truly replace some repetitive labor by humans; but humanoid robots are inevitably the trend of the future, and the winner will be the one who can visualize this "story".