Vol.16 Tsinghua Entrepreneur-Scientist Huazhe Xu: From Intelligent "Hatching" to Home Embodiment, Stubbornly Chasing Originality in the AI Wave

This episode features an in-depth interview with Xu Huazhe, founder of Poko Robotics, the latest embodiment project led by Yunqi Capital's angel round — and something of a special "reunion." Xu sits down with his very first investor, Sang Yu, executive director at Yunqi, for a deep conversation that traces the arc of AI 1.0's technical evolution.

Both Class of 2012 alumni from Tsinghua's Department of Electronic Engineering, they lived through the full sweep from the perception era ignited by AlexNet in 2012 to the pivotal inflection point where embodied intelligence is now scaling the "lofty peaks" toward the home.

Sixty-eight minutes of accessible yet substantive discussion on AI's technical transformation and the path to home robotics — and also a conversation about people and choices, about technical conviction and personal decisions. From the faint glow of a Toronto lab to Berkeley's reinforcement learning boom; from the entrepreneurial thrill of being "fully drained" to the intuitive physics of "ruining a silk blouse." Stories and substance alike — enjoy.

【What You'll Hear】

• Tsinghua EE Reunion: A '90s-born embodied entrepreneur and his investor on the paradigm shifts they witnessed in AI
• Technical Path Deconstruction: Reinforcement learning, world models, online learning — which is the foundational solution to "physical intuition"?
• Defining the Home Robot: How do you build a proactive household member? How do you handle "entropy-increasing" environments and privacy concerns?
• Rejecting "Distillation" Followership: Facing real gaps in compute and data, how does Chinese embodied intelligence build a moat around original intelligence?
• The Scientist-Founder's "Personal Marketization": Is building in public the right choice?
• Commercial Exploration: Beyond the mass-production noise, why do the important things that look "unimportant"?

【Timeline】

PART 1 From 2012 to 2026: The AI Paradigm Shifts We Lived Through

• 01:33 Ten years after graduation: From Tsinghua EE classmates to reunion in embodied intelligence
• 03:18 2012: Back when everyone was still grinding calculus, and neural networks were synonymous with "brute-force overfitting"
• 05:42 Finding deep learning: A Toronto exchange program, stumbling upon neural networks' faint light
• 08:52 Early industry landscape: Andrew Ng's CS229, SenseTime and Megvii beyond Tsinghua's East Gate, and the vanished MRF

PART 2 The Faith in Reinforcement Learning and Seeds Sown in the Valley

• 11:02 The Berkeley moment: How early RL shaped a technical faith
• 17:18 Seeds in the valley: ETH's machine dogs and the underestimated Scaling Law
• 21:08 Choosing to return: That long WeChat Moments post about "the ideal place"

PART 3 Poko Robotics' Technical Choices

• 22:13 Using "failure data" for feedback learning, letting robots learn autonomous trial-and-error
• 24:24 Should embodied world models predict in pixel space or latent space?
• 29:25 Why reinforcement learning? Acquiring "physical intuition" through interaction, like humans do
• 31:48 AGI in digital versus physical worlds — how do the methodologies connect?

PART 4 Why Home Robots?

• 33:39 Product definition: Rejecting fixed-function appliances; it should be a "caregiver-type family member"
• 37:21 Personalization potential: Washing silk blouses or stinky socks? Thousand-person, thousand-face character settings
• 40:03 Why the home track? Dreams, the data breadth of entropy increase, and the ToC commercial picture
• 41:39 Timeline for entering homes: Still far from general-purpose robots, but close to "good products" making it inside
• 44:44 The 2026 inflection point: UMI (body-less) collection and online learning as technical turning points
• 45:52 Solving privacy concerns: Why shorter robot vacuums get accepted more easily

PART 5 Builder Energy and Entrepreneurial Philosophy

• 50:13 Doing self-media for "information equality," and also to take the market's "beating"
• 54:52 Industry "hot takes": Embodied intelligence shouldn't be only about mass production; rejecting low-level "model distillation"
• 59:00 The entrepreneurial experience: Enjoying the feeling of being "fully drained" — please select Difficult mode
• 01:02:04 Self-expectation: Stay original, and obsess over the important things that look "unimportant"
• 01:06:31 Come build with Xu Huazhe — Poko Robotics is hiring!

【Notes: People and topics mentioned in the episode】

• Geoffrey Hinton & Ilya Sutskever: The former hailed as the "Godfather of Deep Learning," co-author of the 2012 AlexNet paper; the latter, former Chief Scientist at OpenAI. Their work launched AI's perception era.
• Pieter Abbeel & Sergey Levine: Titans of robot learning at UC Berkeley, with pioneering contributions at the intersection of deep reinforcement learning and robotics.
• Yann LeCun: Turing Award winner, Chief AI Scientist at Meta. The "latent space prediction" mentioned in the episode is central to the world model route he advocates (JEPA architecture).
• Raquel Urtasun: Founder of autonomous driving company Waabi, professor at University of Toronto, former Chief Scientist at Uber ATG.
• VLA (Vision-Language-Action): Vision-language-action models. End-to-end architectures that map visual perception and language understanding directly to robot action outputs.
• World Models: AI's internal simulation of physical world dynamics. Capable embodied world models should predict future states (e.g., object displacement after force is applied).
• Reinforcement Learning (RL): A mechanism for learning optimal strategies through "trial-and-error" and feedback. The text emphasizes its unique value in processing "failure data" and building "physical intuition."
• PPO (Proximal Policy Optimization): One of the most widely used reinforcement learning algorithms in AI today.
• Online Learning: The ability of a deployed model to continuously self-improve based on real-time interaction data — becoming "smarter with use" rather than locked at factory settings.
• Scaling Law: The principle that model performance grows exponentially with compute, data volume, and parameter scale.
• UMI (乌米): An emerging body-less embodied data collection scheme that can dramatically reduce the cost of acquiring high-quality real-robot data.
• EGO Data: First-person perspective data. Captures human operations through worn devices, providing robots with "human-view" learning material.
• Teleoperation: Human remote control of robot movements to collect motion data during task execution.
• Forward Design: Original R&D from demand goals outward, as opposed to reverse-engineering existing products or simple assembly.

🏢 About Yunqi

Founded in 2014, Yunqi Capital is a focused early lead investor in digital intelligence and hard tech. We have been repeatedly named among China's Top 10 Best Early-Stage Investment Firms by Zero2IPO, ChinaVenture, 36Kr, and other institutions.

Over the past 12 years, we have accompanied 200+ technology startups in their growth, including industry leaders such as MiniMax, JD.com, Manycore Tech, DeepRoute.ai, Neolix, Keenon, Watrix, Astribot, RealMan, Noematrix, PingCAP, Zilliz, Yushikongjian, and XTransfer.

📩 Contact Us For pitch decks or ecosystem partnerships: community@yunqi.vc

WeChat Official Account: Yunqi Capital (ID: yunqipartners)

Xiaohongshu: Anfu Road Yun Xiao Qi

View episode transcript on Xiaoyuzhou