Knowin AI's Fifth Round in 13 Months Takes Its Total Past 1 Billion Yuan for Home Robots

Chinese embodied AI company Knowin AI said on Sept. 29 that it has closed another angel round in the hundreds of millions of yuan, led by a fund affiliated with JD.com, with Loyal Valley Capital, Nanshan Zhanxin Investment and Walden International participating. The Shenzhen-based company has now raised five rounds since it was founded in August 2025, for a cumulative total above 1 billion yuan. Its previous round, about 500 million yuan, was announced on Aug. 10, only 49 days earlier.

Knowin splits its work into two halves. GLOW is a generative embodied foundation model the company describes as the robot's brain; KNOWIN-X1 is the consumer home robot it calls the body. The stated plan is to finish engineering validation and prepare for mass production, with a launch targeted for the first quarter of 2027.

The technical claim the company is pushing is "teach once, learn immediately." A user performs a task a single time and the system encodes the operation into reusable skill information. When the robot faces a new task, the model does not retrain or update parameters; it reads the current scene state and decides the next action on the fly. Demonstrations published by the company show a watering task used to test object generalization, and a box-opening and storage task used to test whether a learned skill transfers when the object, the way it is handled, or its position changes.

In effect, the robot is learning task goals, object relationships and task progress rather than a fixed motion trajectory. That is the difference between a demonstration that works only in the room where it was recorded and a system that can be shipped to homes the company has never seen.

Knowin published benchmark results alongside the round, all of which are company-reported and have not been independently verified: a 62.2% average success rate across 18 simulated tasks in RoboDojo, 86.7% average success on LIBERO-Pro, and 65.62 points for its KnowinBrain-1.5 model on the 2D embodied question-answering track of Embodied Arena. The company says it has more than 200 R&D staff and that more than two-thirds of the algorithm team hold doctorates. It also named Chen Yingcong as an advisor on multimodal models and said Deng Shiyuan has joined as a partner leading infrastructure.

JD.com's involvement is the more revealing detail. The fund is the first major strategic backer from outside the robotics world to lead a Knowin round, and JD.com operates both a large logistics robot fleet and a consumer retail business, plausibly a future channel for a home robot as well as a source of real-world operational data. The company describes its work as spanning generative embodied models, interactive 3D world models and synthetic data, the three ingredients needed to train on scenarios that are expensive or unsafe to collect in homes.

Home robotics has an unusually long gap between demo and product. Tasks that look solved in a video, such as watering a plant or putting a box away, usually hide a long tail of failure cases involving lighting, clutter and objects the model has never encountered. Knowin's bet is that a foundation-model approach which generalizes at inference time closes that gap faster than retraining per task, and that consumers will buy a robot they can teach rather than program.

The first quarter of 2027 is the number to hold the company to. The capital is in place, the model architecture is public and the demonstrations run. What remains is manufacturing a consumer device at a price a household will pay, and shipping one that still works on the day it arrives.