Bringing General-Purpose Robots into Every Home.
KNOWIN Embodied Foundation Model Architecture
To achieve high-level robot autonomy in open-ended environments, we propose a generative embodied AI architecture that establishes a complete closed loop from data and cognition to execution.
1. KnowinDream: A Generative Engine Scaling Past the "Data Wall"
The scarcity of high-quality robotic trajectory data is the fundamental bottleneck constraining the industry. KnowinDream is our data engine built on vision generation models, learning the underlying patterns of scene dynamics from limited real-world data to directly synthesize massive training videos of robotic arm manipulations. This covers extreme lighting, complex textures, and diverse contexts that are nearly impossible to exhaust via manual data collection. Rather than an online component within the inference pipeline, it serves as the bedrock of our entire system's robustness.
2. KnowinBrain: An Embodied Cognitive Foundation with Reflection Capabilities
In the cloud, we deploy KnowinBrain, our foundation VLM (Vision-Language Model) that ranks #1 on the Embodied Arena leaderboard. We innovatively leveraged the massive synthetic data generated by KnowinDream to conduct deep Action CoT (Chain of Thought) post-training, aligning it with Action Token generation to unlock autoregressive action reasoning capabilities. Powered by a 1-5Hz execution feedback mechanism, the post-trained KnowinBrain dynamically evaluates the current physical state, enables autonomous reflection, and corrects policies on the fly, seamlessly closing the loop across perception, planning, and execution.
3. Action Tokenizer: A Generative Action Encoder-Decoder
The Action Tokenizer featured in our architecture diagram is the direct product of pre-training on the massive video datasets synthesized by KnowinDream. Through rigorous training across multi-dimensional physical environments, it masters highly generalized action representation capabilities, efficiently decoding the Action Tokens generated by KnowinBrain into the Action Chunks required by edge systems. This ensures that even when confronting an unfamiliar real world, the robot can still generalizably produce smooth, precise physical manipulations. Furthermore, the Action Token Space is far from static. As Knowin continuously unlocks more complex household application scenarios, this underlying action space will dynamically expand, granting the robot an ever-evolving boundary of physical execution.
KnowinDream
Generative Data Engine
Our generative philosophy extends far beyond the actions themselves. KnowinDream is Knowin's data engine built upon generative models. It learns the underlying laws of scene dynamics from real-world data to directly synthesize training videos for robotic arm manipulations, exhaustively covering extreme edge cases that are practically impossible to capture through manual data collection.
Extreme Scenarios, Single Model
Take folding clothes as an example: the robot successfully operates against a TV screen, dealing with dynamic lighting and constantly changing visual interference. The system was never fine-tuned for any of these specific lighting conditions. This robust zero-shot performance is a direct result of KnowinDream's comprehensive coverage of spatial and visual variations.
About Us
Founded in July 2025, Knowin has pioneered a novel architecture for general-purpose embodied foundation models, defining the next-generation paradigm for Embodied AI. Driven by the mission to bring robots into every home, we are ushering in a new era of human-robot symbiosis. Focused on consumer-grade household scenarios, Knowin develops physically instantiated AI systems capable of continuous interaction, forming a complete closed loop of "Perception - Generation - Execution - Evolution" in the real world.
As a trailblazer in consumer-facing Embodied AI, Knowin leverages full-stack proprietary technology, an entirely new generative embodied foundation model architecture, and an innovative synthetic data pipeline. We are relentlessly advancing our robots' cross-scenario generalization, long-horizon closed-loop execution, and reliable task completion within complex domestic environments. By addressing the diverse needs of modern households, from companionship and interactive entertainment to high-frequency chore assistance, we deliver a smarter, more autonomous, and effortless home service experience.
Contact Us
Phone: 17610340223
Email: HR@knowin.ai
Address: Room A2406, Building 12, Shenzhen Bay Science and Technology Ecological Park, No. 18 Keji South Road, Yuehai Subdistrict, Nanshan District, Shenzhen, Guangdong, China
