China's Homegrown AI Robots Are Getting Wild... Playing Piano, Brewing Tea, Practicing Wing Chun, and Even Petting Cats? | Yunqi Capital

The generalization and versatility of embodied intelligence are starting to show their edge in AI robotics.

Recently, Yunqi Capital portfolio company Astribot officially unveiled its next-generation AI robot, the Astribot S1. Compared to the first technical demo released just three months ago, the S1 has advanced significantly — its mobility and operational range have expanded considerably, the environments and tasks it handles are more complex, and the precision of task execution has reached new heights.

AI technology has opened new frontiers for advancing robotic capabilities. After years of deep engagement in both AI and robotics, Yunqi Capital is now working alongside emerging forces in embodied intelligence to shape a smarter future. Astribot stands as one of Yunqi's representative investments in this space. In this edition of "Yunqi Capital," we take you inside the technical approach behind this innovative company's first-generation AI robot, as well as the founding team behind it.

The following article is republished with permission from QbitAI (WeChat ID: QbitAI). Please contact the original source for reprint requests.

By Mingmin, from Aifeisi Temple

What kind of chops get the CEO of Figure — the hottest AI robotics company overseas — to pay attention immediately?

A Chinese humanoid robot is flexing hard, and its latest tricks include playing the yangqin, brewing gongfu tea... and wok-tossing?! Holding the bamboo mallets, it strikes each string with pinpoint delicacy, producing beautiful music.

It brews a pot of gongfu tea with practiced ease, mastering more than a dozen steps and juggling multiple cups and pots.

It whips up a fragrant waffle, capped off with a signature wok flip.

In its "leisure time," it can even take you through some Wing Chun drills — basically a robotic Ip Man, right?

Note: all of the above actions were performed autonomously by the robot, shown at original speed.

Not only are its hands rock-steady, but every movement flows seamlessly and finishes in one go — smooth like it mainlined a chocolate bar.

Even for something as intricate as gongfu tea, a long-sequence task with over a dozen distinct steps, its "brain" plans everything out clearly. It handles materials ranging from ceramic and metal to wood and tea leaves, and vessels of every shape — bowls, cups, pots, and pitchers — with complete command.

This requires a robot that can learn, plan, and execute like a human, with strong generalization capabilities and general intelligence.

This is the Chinese humanoid robot Astribot S1, officially unveiled recently.

Developed by startup Astribot, it's not just vaporware — it will appear live at the World Robot Conference in Beijing from August 21 to 25 for public demonstrations.

In April of this year, the S1 made its debut with a first technical showcase featuring dozens of difficult yet genuinely useful feats: ironing and folding clothes, sorting and tidying, wok-tossing, vacuuming, and even competitive cup stacking. It sparked significant discussion at home and abroad right out of the gate.

American netizens exclaimed: "China's AGI-level robot is shocking the entire! Industry!"

Less than four months later, Astribot S1 is being officially released in complete form. What does it have in store this time?

Household Chores, Tool Use, Martial Arts — A Decathlon of Skills

A large portion of Astribot S1's skillset covers what people care about most: housework. Its tagline goes:

Go travel worry-free (wander to your heart's content), I'll keep the home warm and waiting. Lounge at home with ease (leave it to me), I'll help you enjoy life to the fullest.

First up: waffles. Warm. The S1 steadily grips a ladle, scoops batter, and pours it into the waffle iron.

After recognizing that the batter has spread across the griddle, it grabs the handle and closes the lid. The robot's waist joint is crucial here. Grabbing the open handle requires a reaching motion. Without the waist coordinating with the upper arms, this kind of extended forward movement would be impossible.

Then it "twists" the knob to heat the batter.

It "grabs" a fork to retrieve the waffle,不忘 showing off a wrist-flip toss.

Next challenge: the looooong process of brewing gongfu tea. Warm. This tests not just fine motor control, but intelligent planning when environment, task, and object complexity all intersect. Tea brewing generally involves rinsing leaves, steeping, pouring, and serving — each stage requiring multiple different tools. The robot must determine which tools to use at each step, and grip a completely different series of items.

It needs to adjust force control in real time. For example, when scooping tea leaves from a ceramic cup, the ceramic is smooth and fragile — too much grip force will damage it. But the very next action is picking up a metal kettle to pour water, so the robot must immediately recalibrate its force control to ensure it can lift the kettle.

Beyond this, Astribot S1 can also use a vacuum cleaner. How does it know to arch its back and drive power from the thighs to reduce lower back strain... must hit the gym regularly.

Feeding the cat — especially warm.

And even remote cat teasing.

The owner just puts on an XR headset and can play with their feline overlord from the office — getting paid to cat-sit!

Astribot S1 captures human hand motion trajectories in real time through the device, then calculates how its own dual arms should coordinate to ensure effective interaction with the cat at home without messing up the room or getting accidentally scratched.

This requires the robot's motion system to respond rapidly and handle complex spatial relationships.

At the same time, Astribot S1 has smoothly replicated even more expert skills.

Like playing the yangqin. Because both strings and bamboo mallets are elastic, the precision force control requirements are extremely high — the robot must provide accurate haptic feedback with every strike and adjust its striking angle in real time. Too light and it won't reach the strings; too hard and it produces noise; too slow and the rhythm falls apart. S1: I'm trying so hard here...

And where last time it threw a paper airplane, this time it's shooting hoops. This action requires full-body posture coordination — starting with a bent waist and raised elbow while holding the ball, then adjusting the wrist as the waist gradually straightens, and finally releasing the basketball with simultaneous power from waist and hand.

There's also the trailer footage of it dancing to "Seaweed" while pulling off an extreme backbend — degrees of freedom off the charts.

Compared to the first technical demo three months ago, Astribot S1 has been on a tear. Its mobility and operational range have expanded, environments and tasks have grown more complex, and precision has reached new heights.

How does Astribot S1 achieve these high-difficulty, long-sequence, generalizable tasks?

High-Value Upper Body + Groundable Lower Body

First, in terms of form factor, Astribot S1 adopts a humanoid upper body + wheeled base approach.

Upper limb manipulation is gradually becoming the focal point of embodied intelligence. In nature, creatures that can use their hands to work are basically highly intelligent mammals, and humans perform most operations with their upper limbs and hands. Therefore, intelligent decision-making and manipulation with the upper body has become the core technical barrier for solving real needs and enabling practical, deployable applications. When the S1 first debuted in April, it positioned itself as having the "strongest manipulation" among robots of its class.

For the lower body, which represents mobility capabilities, Astribot chose a "groundable" wheeled design — high stability, low energy consumption, simple control. More importantly, humans spend most of their time in planar environments: homes, offices, factories, shopping centers. Wheeled bases already cover a large portion of application scenarios.

The head, hands, and torso all use modular designs, allowing flexible assembly or disassembly according to different needs.

The S1's body specifications are extremely human-like. Each arm has 7 degrees of freedom — same as humans. The hand has 2 robotic fingers, capable of completing most tasks, with a dexterous hand also in development.

Its movement and single-arm load capacity both exceed that of an average human male, with positioning precision reaching 0.03mm.

Efficient Multi-Dimensional Data Collection — Breaking the Embodied Intelligence Bottleneck

If a robot's intelligence comes from AI, then AI depends on three elements: data, algorithms, and compute.

GPT's era-defining breakthrough was inseparable from tens of billions of training data points, easily obtainable from the internet. For training robots, directly usable data is essentially zero. Therefore, whether a company can acquire high-quality yet sufficiently cheap data has become the bottleneck constraining robot development today, and a key differentiator in competition.

Astribot holds unique advantages in embodied intelligence data acquisition.

On one hand, the S1 can rapidly learn from massive amounts of real-world video data and human motion capture data. Additionally, the S1 can efficiently collect multi-dimensional, high-quality data — visual, auditory, tactile, and force — from a first-person perspective, closest to real-world experience.

Synthesizing this high-quality data enables Astribot to conduct more efficient large-scale training, while reducing the cost of collecting high-quality robot data, the volume of data needed, and the difficulty of training for new tasks — greatly improving generalization.

As it continues interacting with the real world, the S1 can continuously generate new "learning materials", enabling perpetual learning and evolution toward artificial general intelligence. This is also the key to why the S1 can learn, think, and execute in such human-like ways.

On the AI algorithm front, loading large models gives the S1 perception, cognition, real-time decision-making capabilities in complex environments, as well as intelligent understanding and multi-modal interactive execution abilities — achieving generalizable operation at the object, task, and environment levels.

This means the S1 now has a "super brain," capable of more quickly adapting to new environments and new things, "drawing inferences from one instance," and "grasping one thing and understanding all related things."

Rigid-Flexible Coupled Transmission — Strongest Manipulation, Still Safe

A robot's versatility depends heavily on its body — its hardware or "embodiment."

The S1's "strongest manipulation" comes from its unique rigid-flexible coupled transmission mechanism design. By embedding sensors into the transmission process, the S1 can monitor force transmission in real time. When peeling a cucumber, for example, it doesn't rely on trajectory estimation — like a human, it precisely senses how much force the peeler exerts on the cucumber, then precisely controls its force output. This special transmission structure significantly improves operational precision.

Notably, through rigid-flexible coupled hardware design and innovative force planning algorithms, the S1 achieves extremely high safety. Through this hardware design and force planning, it can precisely control force during interaction — not harming people, objects, or itself during movement. Safety is what makes deployment possible.

Thanks to this meticulously crafted and complete technical solution, every Astribot S1 release manages to deliver that "jaw-dropping" feeling.

The embodied intelligence field is currently in a "hundred flowers blooming" period. Not only are newcomers emerging everywhere, but each company's robot form factors and technical approaches vary widely. How did Astribot S1's technical approach take shape?

From the Hands of Tencent Robotics Lab's First Employee

To answer this, we need to look at the team behind Astribot S1.

Astribot, founded in December 2022, takes its name from the Latin adage "Ad astra per aspera" — "through hardships to the stars" — representing the company's long-term plan and firm commitment to popularizing AI robot technology.

Founder and CEO Lai Jie has 16 years of robotics R&D experience, a veteran in AI and robotics. He joined Baidu in 2014, later serving as head of the "Xiaodu Robot" team.

In 2018, Dr. Zhengyou Zhang, a world-renowned expert in computer vision and robotics and Tencent's highest-level distinguished scientist, was establishing Tencent's RoboticsX lab. Lai Jie joined as employee number one, and later led development of the wheel-legged robot Ollie.

Astribot's other founder, Dai Yuan, earned her bachelor's degree at UIUC and her PhD at UCLA, focusing on robot perception. She has published over 30 papers in top journals including Nature Communications and Science Advances, and holds more than 70 robotics patents. Like Lai Jie, she joined the newly founded RoboticsX in 2018.

Drawing from their experience at Tencent RoboticsX, Astribot places heavy emphasis on "Design for AI" hardware-software integration capabilities, and has built a team structure similar to RoboticsX's — half focused on robot embodiment, half emphasizing using AI algorithms for perception and motion control — to explore deep AI-robotics coupling.

The team largely comes from companies including Tencent, Google, Huawei, and DJI, as well as top domestic and international universities and AI research institutes.

AI Robot Assistants for Billions

From day one, Astribot's vision has been to give billions of people AI robot assistants.

Such an assistant must be able to learn, think, and work like a human; interact with people smoothly and intelligently; use human tools and equipment; and help people complete tedious, difficult, or dangerous tasks — continuously expanding the boundaries of application scenarios and leading the "AI + Robotics" technology revolution.

The company quickly closed tens of millions of dollars in Pre-A funding, led by Matrix Partners China, with follow-on from industrial investors including Daotong Investment and Qinghui Investment, and pro-rata participation from existing investor Yunqi Capital.

From the April technical demo to the August full product launch, Astribot also announced commercialization within 2024. And through self-developed key components, the S1 has clear cost advantages and price competitiveness.

In short, Astribot is moving fast on every front. But perhaps what's fastest is the robot itself. Back in April, founder and CEO Lai Jie put out the call:

"Welcome everyone to make requests of the S1! That's how its capabilities can grow from 55%, to 85%, to 99.99%, infinitely approaching human level."

That Wing Chun demonstration shown earlier? Probably came from a netizen request in April — now realized.

So — what would you want the S1 to do for you?