Skip to main content

RoboScience

Category: Robotics / Embodied AI

A Chinese general-purpose embodied AI company developing the Visics foundation model with VLOA architecture for cross-embodiment dexterous manipulation, enabling robots to autonomously perform complex real-world tasks. RoboScience was founded in 2024. The company is led by Ye Tian (田野). Based in Beijing (HQ), Shenzhen, Suzhou, Hangzhou, China. Team size: 101-500. Total funding raised: $210.0M. Latest round: Series A. Key investors include 京东集团 (JD.com), 零一创投 (01VC), 招商局创投 (China Merchants Venture), 商汤国香资本 (SenseTime Guoxiang Capital), 普华资本 (Puhua Capital), 达晨财智 (Fortune Capital).

Founded
2024
Headquarters
Beijing (HQ), Shenzhen, Suzhou, Hangzhou, China
Team size
101-500
Total funding
$210.0M

Value proposition

Decouples cognitive world models from motor control via Object Trajectory representation (3D point cloud), enabling any robot to manipulate any object to perform any real-world task with zero-shot generalization across embodiments, objects, and tasks.

Products and solutions

Visics General-Purpose Embodied AI Model (VLOA architecture: Vision-Language-Object-Action), RoboMirage Differentiable Physics Engine, Cross-Embodiment Dexterous Manipulation System, Self-developed robot body (upcoming mass production), RobotOS

Unique value

VLOA architecture with Object Trajectory as the intermediate interface — the first to fully decouple 'cognition' (embodied world model) from 'execution' (general manipulation model), trained on massive internet video + simulation data at 1/20 to 1/200 the cost of real-robot data collection.

Target customer

Retail (supermarkets, e-commerce logistics), logistics, healthcare/elderly care, industrial manufacturing, robot OEMs (dexterous hand and robot body manufacturers)

Industries served

Retail, E-commerce Logistics, Healthcare/Elderly Care, Industrial Manufacturing, Commercial Services

Technology advantage

VLOA dual-engine architecture (Embodied World Model + General Manipulation Model); Object Trajectory (3D point cloud) as intermediate representation; RoboMirage high-precision differentiable physics engine; 'Simulation+Video' dual data flywheel achieving weekly hundreds of thousands of hours of training data; Cross-embodiment capability (any robot, any object, any task); ICRA Best Paper awards (D(R,O)Grasp, T(R,O)Grasp with 94.83% success rate SOTA)

How they differentiate

Unlike competitors building hardware-first humanoid robots, RoboScience takes a software-first approach with its VLOA architecture that is inherently cross-embodiment (works across any robot type). The Object Trajectory representation decouples cognition from motor control, enabling generalization without task-specific programming. Their 'Simulation+Video' data flywheel achieves data costs 1/20 to 1/200 of traditional real-robot collection methods.

Main competitors

智元机器人 (Agibot), 千寻智能 (Qianxun AI), 星动纪元 (Xingdong Jiyuan / Robot Era)

Key partnerships

京东集团 (JD.com) - strategic investor and potential deployment partner, 商汤科技 (SenseTime) - via SenseTime Guoxiang Capital, Pilot programs with retail, logistics, and elderly care enterprises

Notable customers

Pilot programs underway with retail, logistics, and elderly care service enterprises (specific names not publicly disclosed)

Major milestones

2024-12: Company founded, 2025-03: Seed round from 01VC, 2025-05: Chief Scientist Shao Lin wins ICRA 2025 Best Paper in Robot Manipulation & Motion, 2025-07: ¥200M Angel round led by JD.com, 2025-09: Launched RoboMirage physics engine, 2025-11: Named to The Information's T150 most promising startups, 2026-02: ¥数百M Pre-A round led by Puhua Capital, 2026-05: ¥1B Series A round, 2026-06: Public unveiling of Visics foundation model with VLOA architecture, 2026-06: Demonstrated furniture assembly (most complex embodied manipulation task), 2026: Forbes China 'AI Technology Breakthrough Enterprise' award

Growth metrics

Raised ~¥1.5B+ (~$210M+) across 4 rounds in ~18 months; 3 rounds of cumulative funding exceeded ¥400M before Series A; Named to The Information T150; ICRA Best Paper awards (Asia's only team to win consecutively)

Market positioning

Premium-tier embodied AI foundation model company competing with top Chinese humanoid/embodied AI startups. Differentiated by software-first, cross-embodiment approach (not tied to specific hardware), VLOA architecture, and world-class research output (ICRA Best Paper awards). Backed by JD.com strategic investment.

Geographic focus

China (primary), with global expansion plans

Patents and IP

VLOA architecture (Vision-Language-Object-Action); Object Trajectory representation using 3D point cloud; RoboMirage differentiable physics engine; D(R,O)Grasp and T(R,O)Grasp cross-embodiment dexterous grasping frameworks (ICRA Best Paper awards)

About Ye Tian (田野)

Ex-Apple AI Platform Tech Lead (7 years, built Apple's ML platform dubbed 'Apple's PyTorch and CUDA'); Stanford AI Lab MS under Andrew Ng; USTC Physics BS (top of class).

Official website: