--- title: "Astribot's bet on hands, not feet: a Shenzhen humanoid that learns by watching" date: 2026-11-06 category: Humanoids & Robotics site: NeuroAI canonical: https://neuroai.site/a/na-robot-astribot-dexterous-humanoid language: en --- # Astribot's bet on hands, not feet: a Shenzhen humanoid that learns by watching > Stardust Intelligence's Astribot S1 hits 10 m/s with its arms and learns by imitation. We verify specs, funding, and why a wheeled torso makes commercial sense. ## Key takeaways - Stardust Intelligence (星尘智能), founded in late 2022 in Shenzhen by ex-Tencent Robotics X lead Lai Jie (来杰), built the Astribot S1 humanoid (人形机器人). - The S1's cable-driven arms hit an end-effector speed of at least 10 m/s with 10 kg payload per arm and ±0.1 mm repeatability — figures confirmed by mainland tech press and a UK robotics directory. - Instead of walking, the S1 is an upper-body dual-arm robot on an omnidirectional wheeled base, trading bipedal flair for manipulation performance. - It learns by imitation: a person demonstrates a task via VR teleoperation, and the robot reproduces it, extended in 2024 by a collaboration with Physical Intelligence's π0 model. - Commercial availability began in limited form in late 2025, with the first units aimed at research customers rather than homes. - Funding includes a 2024 pre-A and an April 2025 Series A backed by Ant Group and Matrix Partners, though exact amounts are not publicly disclosed. ## The video that looked fake In April 2024, a clip circulated that made viewers suspicious for the wrong reason: they assumed it was edited. It was not. A robot arm poured wine without toppling the glass, flipped a pancake mid-air, shaved a cucumber's skin, and wrote calligraphy with a brush — all at a speed that looked unnatural because it was faster than a human hand. The machine was the Astribot S1, and the message was deliberate: for a humanoid, the hands matter more than the feet. That conviction comes from Stardust Intelligence (星尘智能), a Shenzhen startup founded in December 2022 by Lai Jie (来杰), employee No. 1 at Tencent's Robotics X lab. ## The specs, verified The S1 is an upper-body humanoid (人形机器人): two seven-degree-of-freedom arms on an omnidirectional wheeled base. The hard numbers are unusually specific and cross-checked. Mainland outlet Zhidx (under Phoenix) and the UK's robotsbritain.co.uk both report an end-effector speed of at least 10 meters per second, a payload of 10 kg per arm, and positioning repeatability of ±0.1 mm. The speed comes from a cable-driven architecture — drive forces routed through tendons rather than heavy joint-mounted motors — which keeps limb inertia low and makes both fast and delicate moves possible on one arm. The trade-off is explicit. By skipping bipedal balance, the S1 gives up walking for manipulation. The company targets kitchens, labs, service counters, and workcells — flat floors, hands-busy work. It is a different product philosophy from bipedal generalists betting on legs. ## Learning by watching The second thesis is how the robot acquires skills. Rather than coding each task, a human demonstrates it once through a VR teleoperation interface — folding, pouring, assembling — and the S1 studies and repeats the behavior. In November 2024, Stardust extended this with Physical Intelligence's π0 foundation model: given a high-level instruction like "make coffee," the robot decomposed the steps and acted, even when mugs or machines were moved. This is imitation learning riding fast enough hardware to be useful. It is also why the company stresses "Design for AI" — building the body so that data and models, not just mechanics, drive capability. ## Why a wheeled torso, not legs The decision to stay on wheels is a bet on where value is created. Bipedal humanoids spend much of their engineering budget on not falling over; Stardust spent its budget on arms that move faster than a hand and place objects to a tenth of a millimeter. For the environments it targets — kitchens, labs, counters — flat floors are a given and the work is in the grip. That is a narrower ambition than a robot that climbs stairs, but it is also the difference between a demo and a tool. The risk is the opposite: a wheeled upper body cannot follow a worker up a flight of stairs or across a construction site, which caps the addressable market. ## From viral clip to product The S1 debuted publicly at the 2024 World Robot Conference in Beijing that August and moved toward limited commercial availability in late 2025. Notably, the first deliveries target research customers, not consumers — a candid acknowledgment that, in Stardust's own framing, the technology and the embodied-AI field are still tightly coupled and need cooperative iteration. On funding, public records and profiles indicate a 2024 pre-A round with Matrix Partners and Yunqi Capital, followed by an April 2025 Series A with Ant Group, Jinqiu Capital, and Matrix Partners. Exact amounts have not been disclosed. ## Honest limitations - The 10 m/s and ±0.1 mm specs describe peak end-effector performance in controlled demos, not sustained behavior across a full work shift. - "Learning by imitation" works best on tasks similar to demonstrated ones; true open-world generalization remains unproven at scale. - Limited commercial availability in late 2025 means field reliability data from paying customers is still thin. - Undisclosed funding amounts make it impossible to assess valuation or burn rate from public sources. ## What readers can do now - Robotics researchers should evaluate the S1 as a manipulation platform with open APIs and VR teleoperation rather than as a home assistant today. - Buyers should wait for published uptime and task-success data from the 2025 research deployments before projecting it into production lines. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-robot-astribot-dexterous-humanoid Free to quote with attribution and a link to the original.