CtrlK
 | |
| Developer | Astribot (Stardust Intelligence) |
| Type | Upper-body humanoid robot on wheeled base |
| Country of origin | China |
| Unveiled | April 29, 2024 (video); August 21, 2024 (public debut) |
| Height | 170 cm |
| Weight | ~90 kg |
| Degrees of Freedom | 23 total |
| Arm DoF | 7 per arm (14 total) |
| End-Effector Speed | 10 m/s or greater |
| Positioning Repeatability | plus or minus 0.1 mm |
| Payload per Arm | 5 kg (at full horizontal reach) |
| Arm Span | 194 cm |
| Battery Life | 4 to 6 hours active; up to 10 hours standby |
| Locomotion | 3-DoF omnidirectional wheeled base |
| Actuation | Cable-driven |
| Grippers | Parallel-jaw (dual-digit) |
| AI System | Design for AI (DFAI); DuoCore-WB imitation learning |
| Price | $96,000 to $150,000 (estimated) |
| Website | astribot.com |
The Astribot S1 is a mobile dual-arm humanoid robot developed by Stardust Intelligence (Chinese: 星尘智能), a robotics startup headquartered in Shenzhen, China, that operates under the brand name Astribot. The S1 gained worldwide attention on April 29, 2024, when a demonstration video went viral across social media and technology news outlets, showcasing the robot performing household tasks with remarkable speed and precision that surprised observers throughout the robotics community. The robot officially debuted at the 2024 World Robot Conference in Beijing on August 21, 2024, and entered limited commercial availability in late 2025.
Designed as an upper-body humanoid mounted on an omnidirectional wheeled base, the S1 prioritizes dexterous manipulation over bipedal locomotion. Its cable-driven arms achieve end-effector speeds of 10 meters per second or greater with positioning repeatability of plus or minus 0.1 mm, figures that exceed typical human hand performance. The robot learns new tasks through imitation learning, observing human demonstrations via a VR-based teleoperation interface and replicating the demonstrated behaviors autonomously. Stardust Intelligence positions the S1 at the intersection of household assistance, scientific research, and commercial service, with a stated goal of bringing AI robotic assistants to billions of people.
Stardust Intelligence was founded in December 2022 in Shenzhen, Guangdong Province, China. The company's founder and CEO is Lai Jie, who was the first employee (Employee No. 1) at Tencent's Robotics X laboratory. At Tencent, Lai Jie led the development of a wheel-legged robot called Ollie that attracted attention for its ability to perform acrobatic backflips using a tail-like mechanism. Before joining Tencent, Lai Jie served as team leader for Baidu's "Xiaodu Robot" project, an early Chinese conversational robot initiative. He also conducted research at Hong Kong Polytechnic University and holds a master's degree from Wuyi University.[5]
The core founding team consists of six members, all formerly from Tencent's Robotics X laboratory. Beyond this core group, the company has recruited engineers with backgrounds at Baidu, Huawei, Google, and UBTECH Robotics, another Shenzhen-based humanoid robot maker.[5][13]
The name "Astribot" derives from the Latin proverb "Ad astra per aspera," meaning roughly "through hardship to the stars." The parent company name, Stardust Intelligence, carries the same thematic reference in Chinese.[7]
Astribot has raised capital through multiple funding rounds since its founding:
| Round | Date | Key Investors | Notes |
|---|---|---|---|
| Angel | March and October 2023 | Yunqi Partners, Decent Capital | Multiple closings |
| Pre-A | June 2024 | MPCi, Dalton Venture, Qinghui Venture | Described as "tens of millions of dollars" |
| Series A and A+ | April 10, 2025 | Ant Group, Jinqiu Capital, Yunqi Partners, Dalton Venture | "Hundreds of millions of yuan" |
The involvement of Ant Group, the fintech affiliate of Alibaba Group, is notable as it signals interest from one of China's largest technology conglomerates in the humanoid robotics sector. The company has stated that funding is directed toward talent recruitment, research and development, and commercial deployment of its robots.[4][12]
Stardust Intelligence reportedly developed the S1 in approximately one year following the company's founding in late 2022. The robot was designed around a proprietary philosophy called "Design for AI" (DFAI), which co-designs the mechanical hardware and AI software from the ground up rather than developing them independently and integrating them afterward. This approach ensures that the robot's physical structure is purpose-built for AI-driven control.[3][7]
On April 29, 2024, Astribot released a demonstration video that rapidly circulated across social media platforms and technology news outlets. Publications including New Atlas, Interesting Engineering, Fox News, Cybernews, and Tech Times covered the video extensively.[1][2][4] The video showed the S1 performing a series of tasks with speed and smoothness that surprised observers in the robotics community:
When the S1 was first shown in April 2024, it was presented only from the waist up, without revealing its lower body or locomotion system. While the manipulation performance drew favorable comparisons to demonstrations by Boston Dynamics, Figure AI, and other leading humanoid developers, some commentators noted that the absence of a visible lower half raised questions about the platform's completeness. The publication Maginative ran the headline "Stardust Intelligence's Impressive S1 Robot Demo Raises Eyebrows," noting both the remarkable performance and the open questions surrounding the demo conditions.[15] Reader comments on multiple outlets pointed out the heavy use of jump cuts in the video and the presence of pre-placed objects, raising the question of whether the demonstrated capabilities reflected fully autonomous behavior or carefully staged sequences.[1]
The Astribot S1 made its official public debut at the 2024 World Robot Conference (WRC), held at the Beiren Yichuang International Exhibition Center in Beijing's Economic and Technological Development Zone on August 21, 2024. The 2024 WRC featured 27 humanoid robot manufacturers and over 30 upstream and downstream companies from the humanoid robot supply chain, making it the largest gathering of humanoid robot exhibitors in the conference's history.[3][11]
At the event, Astribot showcased the S1's full form for the first time, including its wheeled mobile base, resolving earlier questions about the robot's mobility. The company described the S1 as the "strongest AI robot assistant" and presented it performing a variety of long-sequence tasks at normal speed (not accelerated playback). Demonstrated capabilities included:[3]
CEO Lai Jie stated at the conference: "Our vision is to provide several billion people with AI robotic assistants. Whether it's taking care of the home or working in factories, the more robots learn, decide, and execute like humans, the more they can help people do more and do it better." He noted that the S1's performance at that time reached approximately 55 to 85 percent of human-level competence across different tasks, with the goal of approaching 99.99 percent accuracy over time.[3]
In November 2024, the company demonstrated the S1 making coffee autonomously using Physical Intelligence's pi-zero (pi0) foundation model, a vision-language-action model designed for robotic control. In this demonstration, the robot followed high-level instructions such as "make coffee" by decomposing them into smaller actionable steps: adding water, measuring coffee grounds, and pressing buttons on a coffee machine. The robot's vision system allowed it to identify objects such as mugs and coffee makers even when they were placed in unexpected locations.[16]
For the 2025 Lunar New Year (Year of the Snake) celebrations, Astribot released a video showing its S1 robots preparing for the Spring Festival by cleaning, making tangyuan (glutinous rice balls, a traditional holiday food), and setting off firecrackers. This was part of a broader trend in which Chinese humanoid robot companies participated in the holiday festivities to demonstrate the practical readiness of their platforms.[17]
The following table summarizes the Astribot S1's key specifications, sourced from the company's official product page and the Astribot Suite research paper (arXiv:2507.17141):[6][7][8]
| Category | Specification | Value |
|---|---|---|
| Physical | Height | 170 cm |
| Physical | Weight | Approximately 90 kg |
| Physical | Arm Span | 194 cm |
| Degrees of Freedom | Total | 23 |
| Degrees of Freedom | Per Arm | 7 (14 total for both arms) |
| Degrees of Freedom | Torso | 4 (articulated) |
| Degrees of Freedom | Head | 2 |
| Degrees of Freedom | Mobile Base | 3 (omnidirectional wheels) |
| Manipulation | End-Effector Max Speed | 10 m/s or greater |
| Manipulation | End-Effector Max Acceleration | Approximately 100 m/s squared |
| Manipulation | Positioning Repeatability | Plus or minus 0.1 mm |
| Manipulation | Payload per Arm (horizontal reach) | 5 kg |
| Manipulation | Combined Bimanual Payload | Up to 10 kg |
| Manipulation | Gripper Type | Parallel-jaw (dual-digit) |
| Power | Battery Type | Lithium-ion |
| Power | Battery Life (active use) | 4 to 6 hours |
| Power | Standby Time | Up to 10 hours |
| Power | Charging Time | Approximately 1.5 hours via docking station |
| Mobility | Locomotion Type | Omnidirectional wheeled base |
| Mobility | Maximum Base Speed | 2 m/s |
| Actuation | Type | Cable-driven |
| Connectivity | Interfaces | Wi-Fi (802.11ac), Gigabit Ethernet, Bluetooth, optional 5G |
The company compares the S1's physical capabilities to those of a normal adult male, noting that the robot matches or exceeds human performance in several metrics. Its positioning repeatability of plus or minus 0.1 mm is substantially better than the typical human hand repeatability of plus or minus 1 to 5 mm.[7]
The S1 carries a comprehensive sensor package enabling autonomous navigation, object recognition, and force-controlled manipulation. According to the Astribot Suite paper, the specific sensor hardware includes:[6][8]
| Sensor | Model / Type | Purpose |
|---|---|---|
| Head stereo camera | RGB stereo pair | Visual perception and depth estimation |
| Head depth camera | Orbbec Femto Bolt (RGB-D) | 3D spatial awareness |
| Wrist cameras | Intel RealSense D401 (RGB-D), one per wrist | Close-range manipulation guidance |
| Chest depth camera | Orbbec Gemini 335 (RGB-D) | Mid-range spatial awareness |
| LiDAR | Livox MID-360 | 360-degree spatial mapping and indoor navigation |
| IMU | Inertial Measurement Unit | Orientation tracking and balance |
| Force/torque sensors | At wrists | Contact force measurement for delicate manipulation |
| Tactile pressure sensors | In fingertips | Grip force feedback and slip detection |
| Ultrasonic sensors | Proximity | Close-range obstacle detection |
| Microphone array | Multi-element | Voice interaction and sound localization |
| Temperature sensors | Environmental | Thermal awareness and safety |
All RGB-D cameras operate at 30 Hz. The vision system is compatible with pre-trained computer vision encoders including DINOv2, CLIP, and SigLIP, enabling the robot to leverage large-scale visual pre-training for object recognition.[6]
The Astribot S1 uses a cable-driven actuation system inspired by human musculature. According to the company, this design achieves several advantages over conventional rigid-link robots:[7][8]
The system incorporates what the company describes as a "soft/hard coupling transmission mechanism" with real-time force sensing rather than trajectory estimation, enabling precise control of output force during delicate operations like handling fragile objects or preparing food.[3]
This cable-driven approach stands in contrast to the direct-drive or harmonic-drive actuators used by many competitors. The compliant nature of cable transmission provides an inherent safety advantage: if a person unexpectedly contacts the robot's arm, the cable system naturally absorbs some of the impact force, reducing the risk of injury. The company also states that it self-develops key components, including its high-performance motor drive system, to maintain cost advantages and tighter integration between hardware and software.[7]
Rather than bipedal walking, the S1 uses a three-degree-of-freedom omnidirectional wheeled mobile base with a single articulated "leg" that bends at a knee-like joint. This deliberate engineering decision prioritizes reliability and stability for indoor use cases over the ability to navigate stairs or uneven terrain.[8] The mobile base supports a maximum linear speed of 2 m/s, enabling smooth transit between workstations in household and commercial environments.[6]
This trade-off is common among manipulation-focused robots. Other wheeled-base platforms in the market, such as the Galbot G1, various service robots from Keenon and Pudu Robotics, and upper-body research platforms, make similar design choices to avoid the complexity, power consumption, and fall risk associated with bipedal locomotion.
The S1 is equipped with parallel-jaw grippers rather than dexterous multi-fingered hands. While this limits the range of grasps the robot can perform compared to five-fingered designs (such as those on Figure 02 or the Boston Dynamics Atlas), the parallel-jaw configuration offers high reliability, simpler control, and consistent precision for the household and service tasks the S1 targets. The gripper incorporates tactile pressure sensors in the fingertip surfaces for grip force feedback.[6]
Astribot's core technical philosophy revolves around "Design for AI" (DFAI), a software-hardware integrated architecture that deeply couples AI capabilities with the mechanical system. Rather than developing software and hardware independently and integrating them afterward, DFAI co-designs both from the ground up so that the physical platform is purpose-built for AI-driven control. This approach enables the S1 to learn, plan, and execute tasks in a more tightly integrated manner than would be possible with a generic robot platform running general-purpose AI software.[3][7]
The S1's primary method of skill acquisition is imitation learning. A human operator demonstrates tasks through a whole-body teleoperation interface, and the robot learns to replicate the demonstrated behaviors. In July 2025, Stardust Intelligence published a paper on arXiv titled "Towards Human-level Intelligence via Human-like Whole-Body Manipulation" (arXiv:2507.17141), authored by Guang Gao, Jianan Wang, Jinbo Zuo, and colleagues. The paper introduced the Astribot Suite, a comprehensive robot learning framework for whole-body manipulation, comprising three key contributions:[6]
The teleoperation system uses a Meta Quest 3S VR headset with handheld joysticks, costing under $300 in off-the-shelf hardware. The system supports two control modes:[6]
The control loop operates at 100 Hz with approximately 20 ms end-to-end latency for motor commands. Hand poses from the joysticks map to robot end-effector positions and orientations, which are then converted to joint commands via a whole-body control module.
The DuoCore-WB policy is a transformer-based diffusion model designed to learn whole-body control policies from demonstration data. Key technical details include:[6]
The framework was evaluated on six representative real-world whole-body tasks:[6]
| Task | Success Rate | Skills Tested |
|---|---|---|
| Deliver a drink | 13/15 (87%) | Long-horizon planning, mobile manipulation, articulated objects |
| Store cat food | 19/20 (95%) | Bimanual manipulation in confined spaces, 2 kg payload |
| Throw away trash | 13/30 (43%) | Multi-stage sequential manipulation (lowest rate) |
| Organize shoes | 16/20 (80%) | Whole-body coordination in low-height spaces |
| Throw a toy | 20/20 (100%) | Dynamic throwing movements (peak performance) |
| Pick up toys | 19/20 (95%) | Multi-object sequential manipulation |
Across these tasks, the DuoCore-WB policy achieved an average 80 percent success rate. The authors noted that the cohesive integration of embodiment, teleoperation interface, and learning pipeline represents a significant step toward real-world, general-purpose whole-body robotic manipulation.[6]
The S1 has demonstrated compatibility with external AI foundation models. The November 2024 coffee-making demonstration used Physical Intelligence's pi-zero (pi0) model, a vision-language-action model that allows the robot to accept high-level natural language instructions and decompose them into executable action sequences. Physical Intelligence later open-sourced the pi0 model, enabling broader experimentation across the robotics community.[16][19]
The company has also indicated that the S1 is undergoing integration with large language models for task planning and natural language understanding, which would allow users to issue verbal commands that the robot interprets and translates into manipulation sequences.[8]
The Astribot S1 entered limited commercial availability in late 2025, initially in China, with international rollout beginning across the United States, Germany, Japan, South Korea, the United Kingdom, Canada, France, Australia, and India through 2026. Early commercial pricing falls in the range of $96,000 to $150,000, placing the S1 in the premium tier of humanoid robots.[8]
For comparison with other commercially available humanoid platforms:
| Robot | Approximate Price | Developer |
|---|---|---|
| Unitree R1 | From $5,900 | Unitree Robotics |
| Unitree G1 | From $13,500 | Unitree Robotics |
| Tesla Optimus | $25,000 to $30,000 (projected) | Tesla |
| Astribot S1 | $96,000 to $150,000 | Astribot |
| Figure 02 | ~$100,000 (estimated) | Figure AI |
| Fourier GR-2 | ~$150,000 | Fourier Intelligence |
The S1 is marketed for several distinct use cases:[7][8]
Astribot provides a development toolchain that includes API access, expert development guidelines, a visual development interface, support for major simulation platforms, and AI deployment guidance. This ecosystem is intended to enable third-party developers and researchers to build applications on the S1 platform.[7]
Astribot operates in an increasingly crowded field of humanoid robot developers, particularly among Chinese startups that collectively accounted for nearly 90 percent of global humanoid robot shipments in 2025.[9] The following table provides context on major players as of early 2026:
| Company | Country | Notable Product | Key Differentiator |
|---|---|---|---|
| AgiBot | China (Shanghai) | AgiBot A2 | Industrial and commercial focus; leading global shipment volume (over 5,100 units in 2025) |
| Unitree Robotics | China (Hangzhou) | G1, H1, R1 | Low-cost consumer and research models |
| UBTECH Robotics | China (Shenzhen) | Walker S2 | Automotive factory deployment with BYD and Geely |
| Tesla | United States | Optimus | Integration with Tesla manufacturing ecosystem |
| Figure AI | United States | Figure 02, Figure 03 | Helix VLA model; BMW deployment; BotQ manufacturing facility |
| Astribot | China (Shenzhen) | S1 | Dexterous manipulation speed and precision |
| Fourier Intelligence | China (Shanghai) | GR-2, GR-3 | Healthcare and rehabilitation focus |
| Boston Dynamics | United States | Atlas (electric) | 56 DoF; Hyundai backing; most technically advanced locomotion |
Astribot differentiates itself through its emphasis on manipulation speed and precision rather than unit volume or price competitiveness. While competitors such as Unitree and AgiBot have focused on scaling shipments at lower price points, and companies like UBTECH have targeted industrial deployment on automotive factory floors, Astribot positions the S1 as a premium platform optimized for dexterous task performance in service and household environments. The 10 m/s end-effector speed and plus-or-minus 0.1 mm positioning repeatability represent the top end of publicly claimed manipulation performance among commercial humanoid robots.[8]
The S1's wheeled-base, manipulation-first architecture places it in a distinct subcategory of humanoid robots that sacrifice bipedal locomotion for superior arm dexterity and reliability. This approach contrasts with fully bipedal platforms:
| Feature | Astribot S1 (Wheeled) | Bipedal Humanoids (e.g., Figure 02, Optimus) |
|---|---|---|
| Locomotion | Omnidirectional wheels; flat surfaces only | Bipedal walking; can navigate stairs and uneven terrain |
| Stability | Very high; no fall risk | Requires active balance control; fall risk exists |
| Manipulation Focus | Primary design priority | Often secondary to locomotion development |
| Power Efficiency | Higher (no energy spent on balance) | Lower (constant balance computation and actuation) |
| Indoor Suitability | Excellent | Good, but overkill for flat environments |
| Outdoor Capability | Very limited | Better suited for varied terrain |
| Mechanical Complexity | Lower (fewer moving parts in base) | Higher (leg joints, ankles, hip actuators) |
Astribot's development takes place within a broader national push by China to establish leadership in humanoid robotics and embodied AI. Several government initiatives at the central and local levels support this effort:[10][14]
This policy environment provides favorable conditions for startups like Astribot, offering access to subsidized facilities, government procurement opportunities, and a growing domestic market. Industry analysts have noted that Chinese humanoid robot companies benefit not only from direct government funding but also from the country's mature electronics supply chain, which provides low-cost sensors, actuators, and computing hardware.[10][14]
Despite its impressive demonstrations, the Astribot S1 faces several acknowledged limitations: