Embodied AI Explosion on the Horizon. Liang Wenfeng Intensively Engages with Startups, DeepSeek May Enter Robotics Directly

In the autumn of 2026, a 'steel wind' swept through the AI circle — not the cold glow of chips, nor the computational roar of large models, but the sensation of realness that comes with the turning of metal joints, the slight heating of sensors, and the moment when a robotic arm first steadily takes a cup of coffee handed by a human. It's that heart-pounding sense of authenticity.

On October 13th, Kuaikjia (Fast Tech) exclusively revealed: Liang Wenfeng, founder of DeepSeek, has recently transformed into a 'embodied intelligence hunter,' meeting over seven hard tech startup teams for two consecutive weeks. His travels spanned Shenzhen Nanshan, Hangzhou Future Science and Technology City, Shanghai Zhangjiang, and Beijing Zhongguancun. What was discussed wasn't PPT parameters, but whether motor response latency could be pressed under 8 milliseconds, whether multimodal VLA models could achieve real-time relocalization during bumpy movement, and — when a large model suddenly 'got it' and wanted to autonomously twist open a bottle cap, would the robot body dare to execute?

Exactly, this is the hottest sector right now: embodied intelligence (Embodied AI). It's not about letting AI 'talk about ideals' in servers, but giving the brain a body, allowing algorithms to learn to walk on legs, grasp objects with hands, and understand the 3D world with eyes — in short, it's the physical gateway to general artificial intelligence (AGI). And in 2026, this gateway is accelerating from the lab toward the production tipping point.

Why has embodied intelligence suddenly become the top trend?

Looking back at 2025, Tesla's Optimus Gen-2 achieved a breakthrough in end-to-end grasping task success rate exceeding 91%; Boston Dynamics' Atlas completed full autonomous warehouse inspection + anomaly identification loop; domestic Unitree Robotics' Go2 quadruped robot has connected to DeepSeek-V3 large model, and in real-world testing at a Dongguan electronics factory, it autonomously disassembled faulty industrial control cabinets and generated maintenance reports... These are no longer demo videos, but actual B-end orders being delivered. According to IDC's latest '2026 Global Embodied Intelligence Industry White Paper' forecast: the size of the Chinese embodied intelligence hardware market will reach 89,000,000,000 yuan, with a year-on-year growth rate of 67.3%, far exceeding the growth rate of the large model software market — because the market has finally realized: an AI without a 'body' is like a student who hasn't passed the driving license exam but proudly claims to be able to drive a rocket, no matter how smart, it can't go anywhere.

Li Wangfeng's 'Three Moves': Investment, Joint Research, and Self-Construction?

According to insiders, Liang Wenfeng's recent intensive contact was not just polite small talk, but rather with a clear 'Embodied Intelligence Combat Map':

The first move has already been made — heavy investment in Unitree Robotics. In August of this year, DeepSeek strategically invested 141,000,000 yuan and signed a tripartite 'Strategic Cooperation Memorandum of Understanding'. Note that this is not a regular financial investment, but a 'capability binding': DeepSeek-V3 large model will be deeply adapted to the underlying OS of Unitree's H1 humanoid robot; both parties will jointly build an 'Embodied Reasoning Engine' (Embodied Reasoning Engine), focusing on compressing the 'perception-planning-action' pipeline; more importantly, Unitree will open all motion control API interfaces, allowing DeepSeek engineers to directly call torque commands — this is equivalent to handing the robot's 'cerebellum' over to AI for fine-tuning.

The second move is in progress — leading investment in WinverseAI. This company, founded by former Tencent AI scientist Hu Han, appears to be doing model inference acceleration chips on the surface, but its core ambition is huge: its self-developed 'NeuroMotion Core' architecture allows 10B-level VLA models to achieve 20FPS real-time spatial understanding at the edge. In short, it allows robots to make decisions on the fly without needing to 'send data back to the cloud for a long time of thinking'. Although some in the industry have jokingly remarked that 'WinverseAI isn't a pure embodied company, more like a 'turbocharger' for embodied intelligence', it's precisely this kind of foundational infrastructure that has been holding the entire industry back. Currently, Meituan Longzhu Capital and Red Sea China Phase II Fund have both confirmed follow-on investments, with a post-investment valuation of 400,000,000 USD — and if DeepSeek leads the investment successfully, it will become the first Chinese AI company to connect the 'large model + embodied OS + edge inference chip' full-stack chain.

The third move is yet to be revealed — personally stepping in to build a team. Several technical leaders close to Liang Wenfeng have revealed that an internal project, codenamed 'Project Atlas', has already been launched: two 'Embodied Intelligence Laboratories' are being established simultaneously in Beijing Yizhuang and Hangzhou Binjiang, with core directions targeting 'low-cost, high robustness dexterous hands' and 'long-term task scheduling systems driven by large models'. One employee joked: 'Liang Total has even moved his office to a floor close to the robot testing field, and every day before leaving work, he must go to watch the robotic arm stack LEGO blocks.'

Embodied Intelligence ≠ Robot + AI, but rather a 'Cognitive Revolution'.

Here, we must emphasize: Many people mistakenly understand embodied intelligence as simply 'a robot with a large model', which is completely wrong! True embodied intelligence is about allowing AI to reconstruct its cognitive framework through continuous interaction with the physical world — just like an infant grasping, falling, and getting up again, each touch trains its 'world model'. If DeepSeek truly enters this field, its biggest advantage may not lie in hardware, but in its so-called 'neural-symbolic hybrid training paradigm' (NS-Hybrid Training): using symbolic logic to constrain large model hallucinations, and using reinforcement learning to cultivate physical intuition, ultimately producing not just an 'AI that can answer questions', but an 'embodied agent' that understands limits, knows when to act, and can make balanced decisions.

.globally broader battlefield: from factories to homes, embodied intelligence is breaking through barriers

In 2026, the application scenarios of embodied intelligence have already moved beyond the stage of laboratory showmanship: • At the BYD Xi'an battery plant, the joint deployment of inspection robots equipped with Unitree and DeepSeek can not only identify cracks in battery cells as small as 0.1mm, but can also autonomously trigger the fire sprinkler system in case of sudden smoke; • In a certain elderly community in Hangzhou, service robots equipped with a lightweight DeepSeek-Edge model have already achieved full-chain services including 'recognizing dialects → identifying emotions → adjusting medication boxes → accompanying conversations', where when an elderly person says 'my stomach is not feeling well today', the robot immediately retrieves the medication records from the past three days and contacts medical staff; • Even in the education sector, 'AI embodied teaching tools' have entered classrooms in primary and secondary schools in Shenzhen — students program robots to complete lever experiments, and the system provides real-time feedback on torque deviation, making knowledge tangible and resistant.

.conclusion: When Liang Wenfeng walks into the robot laboratory, he is signing not just an investment agreement, but an invitation letter for an 'AI coming-of-age ceremony' —

inviting AI to step out of the data cage and enter the real world of wind, rain, sunshine and light; inviting China's AI industry to upgrade from a 'computing power arms race' to a 'physical world operating system' battle; and more importantly, inviting every ordinary person to think: in the next decade, will we coexist with machines that are becoming more and more human-like, or will we coexist with machines that are becoming more and more understanding of us?

Embodied intelligence is not a future technology, but something happening now. And DeepSeek's recent move may just be the key leap for China's AI to transition from a 'follower narrative' to a 'defining moment'.

(Note: As of the time of publication, DeepSeek's official has not commented on its layout in embodied intelligence. However, all clues come from cross-verification by frontline entrepreneurs, supply chain partners, and several industry insiders who participated in closed-door technical salons. We choose to believe: when top large model companies start seriously studying 'how to prevent robots from spilling coffee', the spring of embodied intelligence has truly arrived.)

#embodied intelligence #Deep,Seek #Liang Wenfeng #AI robot #2026 tech trends #embodied intelligence Return to the previous page for more information