Challenging Boston Dynamics, humanoid robots are causing a stir in Silicon Valley.
K-scale roughly fits all your imaginings of a Silicon Valley startup. Under the dazzling North California sun, a group of young engineers are gathered in a garage. With tousled hair and sleepy eyes, their fingers fly over the keyboards, and lines of code scroll rapidly on the screens.
Beside the computer desks, several empty beer cans left over from last night's team party lie haphazardly. Around them are randomly placed robotic arms, a continuously working 3D printer, and various entangled data and charging cables. The air seems to be filled with the mixed smell of solder, coffee, and dreams.
Suddenly, an engineer's slightly hoarse but extremely excited voice cuts through the sound of keyboard typing: "Are you ready? Countdown: three, two, one!"
Everyone's eyes turn in unison to a robot in the middle of the garage that was previously silent and listless. As if an invisible command has activated its core, the indicator lights on the robot suddenly light up. First, it stiffly moves its "arms", then tentatively stretches its "legs". In the breath - held anticipation, it staggers but firmly takes the first step that will be recorded in the team's log. Although its posture is still inexperienced, this "small step" is a giant leap towards the future for K-scale Lab.
This process, which seems as exciting as a rocket launch to outsiders, is called "Deploy" by the engineers. The essence of it is to "pour" the robot's movements and decision - making scenarios that have been simulated and iterated countless times through the reinforcement learning algorithm in software into the physical robot, enabling it to reproduce almost the same actions in the complex and ever - changing real physical environment as in the software simulation.
K-scale Lab, a new and rising force in the Silicon Valley robotics field that has been established for less than a year, is moving towards their grand vision in such repeated "deployments" and iterations: to build a truly open - source and inclusive robotics platform that integrates hardware and software.
A Firm Supporter of the Reinforcement Learning Approach
In the air of Silicon Valley, the wave of artificial intelligence is sweeping everything with unprecedented intensity. However, for Xu Rui, the co - founder and chief operating officer of K-scale Lab, the future is not just about algorithms in the cloud. What he imagines and strives for is an embodied intelligent agent that can truly move in the physical world and continuously evolve with the "clumsiness" and "intelligence" similar to a child learning to walk. What drives all this is his and the team's almost all - in belief in "Reinforcement Learning, RL".
"We use reinforcement learning one hundred percent to do the entire motion control of the robot," Xu Rui says with unwavering determination. "There are nine people in our team now, and most of them are engineers specializing in reinforcement learning. I'm the only one in the company who doesn't write code," Xu Rui says with a smile. "But it also means that I'm responsible for all things outside of technology and engineering."
This is not just a choice of technical path; it's more like an adherence to an underlying philosophy. Globally, many humanoid robot companies still use the traditional motion planning method, presetting every movement of the robot through precise mathematical models. However, the team of K-scale Lab has chosen a completely different and less - traveled path.
"We don't do any traditional motion control; it's all about reinforcement learning," he explains. Traditional methods often struggle and lack stability when facing dynamic and changing environments, such as a sudden gust of wind or an unexpected bump on the ground. Reinforcement learning, on the other hand, is expected to achieve stronger generalization ability.
The concept of this method sounds somewhat like the principle of "the simplest truth is the most profound", but it is extremely difficult to implement. Imagine teaching a baby to walk. You don't write instructions for every muscle contraction but set goals: don't fall, walk forward, and alternate the left and right legs. What the K-scale Lab team does is similar - defining a "Reward Function" for the robot. "For example, we set that the robot cannot fall, its legs must alternate forward, it cannot jump in place, and it must maintain a certain speed while moving forward." The robot then figures out how to maximize these "rewards" through repeated falls and getting up in the simulated environment, and finally "learns" to walk.
The choice of reinforcement learning also profoundly affects their hardware selection.
"I don't think the hydraulic system will be the future direction because it can't support reinforcement learning well," Xu Rui's words imply a subtle challenge to industry predecessors like Boston Dynamics, which are well - known for their hydraulic - driven robots.
In his view, electric motors, especially those joint motors that benefit from the progress of electric vehicle technology, are a better choice at present.
"Why has the humanoid robot become popular again? Isn't it because of reinforcement learning?" he asks rhetorically, pointing out the core driving force of technological iteration.
As many people have judged, the rise of robot startups is promoted by three factors: large - language models (LLMs) and multi - modal models based on the Transformer architecture have demonstrated the amazing potential of general artificial intelligence to the world and provided the possibility of stronger decision - making ability for robots. Second, the maturity of the hardware industry chain. Thanks to the rapid development of the electric vehicle industry chain, the costs of high - performance servo motors, sensors, and battery technologies have been continuously decreasing, laying the foundation for manufacturing more flexible and efficient robots. Finally, the eagerness of capital. After witnessing the great success of generative artificial intelligence represented by OpenAI, capital is shifting from pure general AI algorithms to "hardcore technology" that can deeply interact with the physical world.
The "Geek" Vanguard: Bringing Humanoid Robots into Ordinary Enthusiasts' Homes
Who will be the first users of such an advanced robot that almost completely relies on "self - learning"? Surprisingly, it's not the factory assembly lines, high - end research institutions, or ordinary household consumers as people usually imagine. The first large - scale humanoid robot carefully developed by the K-scale Lab team has 24 degrees of freedom and is priced attractively at less than $10,000. Its target users are those enthusiastic "Geeks" in the US market.
"We hope to sell this product to real enthusiasts first, so that they can take it home and tinker with it," Xu Rui describes his target users. He hopes that after getting the robot, these geeks can program and train it by themselves, just like they would with an open - source computer, and make the robot learn to tidy up the room, dance a funny dance, or even complete more imaginative customized functions.
This is not just a clever market entry strategy; it's also a well - thought - out data accumulation plan. Xu Rui admits that the biggest bottleneck of reinforcement learning at present is the "lack of data volume".
"The ceiling of RL now is data. There is simply too little effective interaction data for robots," he says. Different from large - language models that can consume all the text on the Internet, the learning of robots relies more on interaction data from the real physical world. By first putting the robots into the hands of the most creative and enthusiastic enthusiasts, they hope to generate a large amount of diverse usage scenarios and data to feed back into the model's iteration.
In addition to this "big guy", a small - scale robot priced at about $1,000 is also in the works. Its potential customer group is high - school parents who hope that their children can enhance their competitiveness in college applications by participating in cutting - edge technology projects. The market enthusiasm is beyond imagination. According to Xu Rui, this small - scale robot has already received up to 20,000 purchase intention registrations.
Although initially focused on enthusiasts, Xu Rui has a more long - term vision. He believes that as technology matures and data accumulates, the product's functions will gradually become more general and penetrate into more practical scenarios, such as simple dictation and cleaning. At the same time, they also pay attention to the industry and school markets, believing that there is still a lot of room for improvement in the software openness and technical support of existing educational robots.
The Adherence to Hardware - Software Integration and the Soul of the Open - Source Community: The Birth of ksim
"There aren't many companies in the US that do both hardware and software and integrate them deeply like we do. Many only focus on software or AI training," Xu Rui points out one of the team's core competitiveness. They firmly believe that the physical performance of hardware and the intelligent algorithms of software must be deeply coupled to unleash the maximum potential of humanoid robots. In hardware development, they have three core evaluation criteria: usability (able to smoothly perform basic actions), robustness (durable and not easily damaged), and cost - effectiveness (affordable).
Xu Rui has rich work experience in several large technology companies in China and the US, especially in the field of intelligent hardware. This experience has given him a deeper understanding of the characteristics of information density and innovation environment in China and the US, believing that "these two places are currently the most information - dense and suitable for robot and AI development in the world."
In terms of manufacturing, the team initially planned to rely on China's mature supply chain system. For example, the dexterous hands of the robot use mature products from domestic suppliers.
However, due to complex geopolitical factors, they have to plan ahead and consider the possibility of moving the production line to Southeast Asia or even building a factory in the US.
"I think the talent reserve in the US manufacturing industry is indeed a problem," Xu Rui says with concern. "Many experienced craftsmen are getting old, and there is a gap in manufacturing experience among the younger generation." He even half - jokingly sighs: "The 'dream' (referring to cost) in the US isn't low enough!"
What better demonstrates K-scale Lab's determination to be open than hardware manufacturing itself is their extreme pursuit of "open - source". They not only have an open - minded concept but also put it into practice. They publicly released an open - source Python library called 'ksim', which has been updated and iterated more than 60 times. This powerful tool is built on Google's MuJoCo physics engine and JAX high - performance computing library, aiming to provide a convenient and efficient robot testing platform for the global developer community.
Through ksim, developers can easily conduct reinforcement learning simulation training, "running in the simulator again and again until satisfied", continuously iterating and optimizing the algorithm, and finally "deploying" the trained model with one click to the real K-scale robot for rigorous testing and verification in real - world scenarios. This not only greatly lowers the threshold for robot AI development but also truly reflects their open - source spirit of "hoping that everyone can participate".
The community is another important cornerstone in the company's blueprint. On the social platform Discord, they have gathered an active community of thousands of people with diverse backgrounds, including senior enthusiasts, college students, and curious teachers. These community members are not only potential users of the product but also active testers and co - builders, exploring various possibilities with tools like ksim. Some even use 3D printing and other tools to make their own robot hardware to match the team's open - source software.
An employee from Alibaba managed to successfully replicate their robot design just by studying the publicly available project information, which both surprised and pleased the K-scale Lab team.
This complete openness and active community operation are regarded by Xu Rui as the "moat" that differentiates them from other robot companies, especially those domestic projects that "claim to be open - source but actually have core technologies and components that are difficult to replicate".
"Many domestic companies claim to be open - source, but to be honest, their so - called 'open - source' is something that no one else in the world can fully replicate except themselves," Xu Rui says.
Of course, choosing open - source also means having to bear its inherent "uncertainty". "We've been 'delaying' the product since last November," Xu Rui explains frankly. "The reason for the delay is simply that all our toolchains are open - source, and the biggest problem with open - source toolchains is that they are sometimes unreliable." However, this adherence to openness stems from their firm belief in the power of the community and long - term value.
In terms of sensor solutions, they also follow the RL - first approach, mainly relying on vision and IMU (Inertial Measurement Unit) instead of LiDAR, which is widely used in the robot field.
"Our RL algorithm is based on vision and gravity sensors. Even if you install LiDAR, it won't be useful for our system," he says. This is not only a cost - saving consideration but also in line with the end - to - end learning concept advocated by companies like Tesla.
The Young "Special Forces", Tesla Genes, and the Changing Wind Direction of Capital
What drives all this is a lean team of only nine people. Most of the members are experienced reinforcement learning engineers, full of the vitality and sharpness of youth.
More notably, Xu Rui reveals that another co - founder of the team was a core member of Tesla's FSD (Full Self - Driving) V12 version.
"If you say you're in the AI field now, investors might just say 'oh, okay'; but if you say you're entering the manufacturing or hard - tech field, all investors' eyes will light up, and they'll be eager to invest in you right away!" The focus of capital seems to be shifting from pure general AI algorithms to "hardcore technology" that can deeply interact with the physical world.
The company has successfully completed its angel - round financing, with an amount of less than $5 million. The investors are mainly from the US. The team - building method is also very Silicon Valley - style. Many members are recruited from the active community, first working as contractors on short - term contracts. Their abilities and compatibility are evaluated in real projects, and those who perform well are then converted into full - time employees.
Xu Rui jokes that this is similar to the rise of the Internet industry: "In the current robot and AI fields, most of the capable ones are also young people."
The choice of K-scale Lab's current "headquarters" is also full of Silicon Valley's legend and practicality. This elite team of less than 10 people has not chosen a traditional office park but rented a five - bedroom single - family mansion in Atherton. Atherton is a name that represents the top location in Silicon Valley and even the whole of the US, always topping the list of the most expensive zip codes.
They pay a monthly rent of about $20,000 for this mansion, which is valued at about $15 million in the market. The owner of the house is rumored to be a former vice - president of PayPal.
It may sound incredibly "extravagant", but in Silicon Valley, where space is at a premium and startup costs are high, it's a well - thought - out and "cost - effective" decision for a startup team like K-scale Lab.
The founder team and early employees have completely integrated their work and life. The downstairs living room and garage have become a "battlefield" and laboratory filled with circuit boards, 3D printers, and robot prototypes, while the upstairs bedrooms and the spacious backyard are their "harbors" for short breaks. This arrangement not only saves a significant amount of money on independent office space rental but also, more importantly, greatly improves communication efficiency and work output as team members live, eat, and work together almost 24/7, whether it's a sudden brainstorming session at the dinner table or solving technical problems together in the laboratory late at night.
Located in the heart of Silicon Valley, there is also a unique advantage: there are plenty of relevant talents around. Xu Rui says that K-scale is only a few minutes' drive from Stanford University. Whenever they encounter a problem, they can easily invite a few Stanford students on a weekend, and the problem will be solved quickly.
The Inspiration from Tesla, the Strength of Chinese Hardware, and the Dilemma of "Pseudo - Open - Source"
As an insider, Xu Rui has his own unique observations on the current situation and future of the robot industry. He believes that Tesla's FSD V13 version has made "great progress" in the end - to - end reinforcement learning approach, and its "smooth" experience convincingly proves the great potential of the RL approach. Although he also admits that RL "definitely has a ceiling", he adds, "At least for now, we haven't seen where that ceiling is."
He has reservations about traditional sensor approaches like LiDAR, believing that they are costly and may not be a long - term solution. When talking about the global competition pattern, he admits that the level of Chinese hardware in the humanoid robot field is "quite impressive". Taking companies like Unitree as an example, "they may not be inferior to Boston Dynamics." This recognition of Chinese hardware strength is also one of the reasons why they hope to find