S1

It went viral in 2024 for a demo of fast, fluid household manipulation (pouring wine, flipping sandwiches, calligraphy, Wing Chun). Physically it is not a walking humanoid: a 3-DoF omnidirectional wheeled base (max ~2 m/s) carries a 4-DoF articulated torso that squats and reaches from floor to ~2 m, a 2-DoF head, and two 7-DoF cable-driven arms ending in parallel-jaw grippers, for 23 total DoF. Each arm holds up to 5 kg and moves at end-effector speeds of ≥10 m/s with ±0.1 mm repeatability; runtime is 4 to 6 hours. Skills are taught by VR teleoperation and imitation learning (Astribot's DuoCore-WB diffusion policy), reaching ~80% success on demonstrated tasks, and driven through a Python/ROS 2 SDK. Commercially, Astribot became a unicorn (over 10B RMB valuation, Ant Group among backers) and began delivering thousands of its cable-driven units at the end of 2025 through industrial and commercial cooperation orders. That makes the S1 real and deployed, but B2B-only: there is no public price and no consumer or self-serve order path, and the many "buy online" listings are unauthorized aggregator pages.

Day one with zero programming it does essentially nothing useful autonomously: the S1 is a manipulation research/deployment platform whose skills come from human demonstrations. An operator can teleoperate it via a VR headset and handheld controllers to perform bimanual tasks in real time, and any pre-trained demo tasks the vendor provides can be replayed, but there is no built-in library of household or industrial chores that runs on its own without task-specific data collection and policy training.

Best for: Robotics and embodied-AI teams, manipulation research labs, and industrial/commercial integrators that want a fast, precise bimanual manipulation platform they will teach via teleoperation and imitation learning, and who can engage a Chinese manufacturer directly on an enterprise basis.

Robotics/ML engineering. The intended workflow is VR-teleoperation imitation learning: an operator demonstrates a task (typically ~10 to 50 demonstrations), the trajectories are recorded, and a visuomotor policy (Astribot's DuoCore-WB diffusion policy) is trained, reaching roughly 80% success on demonstrated tasks. Teams work through a Python SDK with ROS 2 integration, high-level task primitives plus low-level joint/Cartesian control, and camera streams exposed as ROS topics. You need people comfortable with data collection, policy training, GPU compute, and integrating the robot into your environment. A wheeled dual-arm robot of ~80 to 90 kg also needs space, charging, and safety planning around people.

SDK access: Astribot exposes a Python SDK with ROS 2 integration: high-level task primitives as Python calls and lower-level joint/Cartesian control for custom controllers, plus onboard-camera vision exposed as ROS topics. The official product page notes S1 specifications "can unlock to higher performance levels" for large research/developer/scenario customers, implying tiered/negotiated capability access rather than a single public developer SKU. SDK terms, access, and any EDU/developer program are arranged through Astribot directly and are not publicly documented with pricing.

Price and availability verified 2026-10. Spotted a change? Tell us and we fix the listing.