Visincept - 视启未来

Visincept
视启未来

Shenzhen IDEA spin-off building vision models and world models for robots; its WorldIncept model topped the WorldArena 2.0 and TriWorldBench boards in 2026.

Visit Website

Company Overview

Visincept is a Shenzhen computer-vision and physical-AI company incubated by the Computer Vision and Robotics (CVR) team of the International Digital Economy Academy (IDEA). The operating entity is registered in Shenzhen, Guangdong. Its founder and chief executive is Dr. Zhang Lei, a chair scientist at IDEA. The company's stated aim is to build vision-native world models that let robots perceive, reason about and predict the physical world, and it describes its core business as private model deployment, embodied-AI vision solutions and model-as-a-service access through the DINO-X Platform.

The team is known for a line of open-set detection research published at CVPR, ICCV, ECCV and ICLR, including DAB-DETR, DN-DETR, DINO, Mask DINO, Grounding DINO, Grounded SAM and T-Rex2. The commercial model family built on that work is DINO-X, a unified vision model for open-world object detection, segmentation, keypoints, captioning and OCR, offered with a cloud playground and API, plus Grounding DINO 1.6 Pro, Grounding DINO 1.6 Edge for on-device use, DINO-XSeek and the T-Rex2 visual-prompt detector. DINO-X Video extends this to cross-frame event understanding for security and retail video.

The robotics line has three parts. DINO-X Grasp is a universal grasping model that uses open-set detection plus 3D depth and shape perception to drive robot arms through perceive, recognise, plan and execute steps; the company lists home tidying, parcel sorting, food packaging and waste sorting as target scenarios and marks the model as coming soon. The DINO-X World Model, which the company enters in benchmarks as WorldIncept, is a latent-space world model for generation and understanding of robot manipulation video. EgoTwin, launched with Baidu AI Cloud in May 2026 at the Global Terminal AI Expo in Shenzhen, reconstructs 3D hand keypoints and hand meshes from first-person video and aligns them to a robot state space, as a cheaper route to manipulation training data than teleoperation.

WorldIncept placed first overall in the WorldArena 2.0 world-model challenge (run with an IROS 2026 workshop) and, as of 30 October 2026, held first place with a score of 71.84 out of nearly 70 models on TriWorldBench, a multi-view bimanual manipulation benchmark built on RoboTwin 2.0 by Peking University, Tsinghua University and other institutions.

Alongside the models, Visincept sells applied tools: T-Rex Label, an annotation accelerator; CountAnything, a counting application; and the open-source DDS (Deep Data Space) tool for dataset visualisation and model analysis.

The company closed an angel round of nearly RMB 100 million in 2025, reported as led by the A-share listed chip maker Anyka (Ankai) Microelectronics with participation from several investment firms. Headcount and founding date are not published.

Capabilities & Activities

Primary type & automation activities this supplier delivers:

Applications & Industries

Product Categories

Products & Solutions

DINO-X

Open-world vision model for detection, segmentation, keypoints, captioning and OCR, offered via API and playground.

DINO-X Grasp

Universal grasping model that drives robot arms to detect, localise and pick arbitrary objects in 3D.

DINO-X World Model (WorldIncept)

Latent-space world model that predicts multi-view robot manipulation video for robot learning and simulation.

EgoTwin

Data engine that reconstructs 3D hands from first-person video and aligns them to robot state space.

Partnership & Notable clients

Associating Events

Listed in RobotToday Supplier Discovery. Verified Profile

Contact Visincept

WEBSITE

https://visincept.com/

EMAIL

[email protected]

HEADQUARTERS

Shenzhen

China

Company Facts

Founded

-

Primary Role

Software/Algorithm

Company Size

-

Primary Region

Asia-Pacific

Annual Sales

-

Funding Stage

Early-Stage VC

Funding Total

-