沈宇军 | 张小珺Jùn｜商业访谈录 Summary

1. Key Themes

The "Embodied-Native" Bet: Abandoning Digital-World Models Entirely

Shen Yujun and Lingbo made the decisive bet to retrain all vision-related models from scratch for the physical world, rejecting the shortcut of fine-tuning digital-world models. The core insight is that digital-world models optimize for the wrong objective — semantic understanding and visual quality — rather than spatial perception and real-time action.

"In 26 this year, or rather this half-year, we made what I think is a fairly big decision: since we can't push the digital-world teams to change, we'll redo everything from scratch based on the physical world's needs." 00:22:24

"Although architecturally we've already achieved embodied-native, I feel the data side hasn't caught up yet." 00:09:00

The Semantic vs. Spatial Gap: Why Digital Vision Models Fail Robots

Shen articulates a precise, non-obvious distinction: semantic understanding (knowing "there's a cat") is fundamentally different from spatial perception (knowing the cat is behind a glass door you'll walk into). This distinction drives Lingbo's entire vision pre-training redesign.

"You can see it but you can't touch it. Seeing it is semantic understanding; not being able to touch it is spatial perception... If you're walking forward you'll hit the glass first, then there's the cat. So we'd try to slide the glass door open, and as it opens you see the cat gradually appear from behind the door. But from a semantic understanding perspective, the cat was always there." 00:45:15

"We built our pre-training purely from a geometric angle — you can roughly understand it as points, lines, and surfaces. Starting from edges, we let the model better understand the concept of geometry." 00:42:39

Data Scale-Up Is the True Bottleneck to Robot GPT-1

Despite architectural progress, the GPT-1 moment for robotics hasn't arrived because robot-specific data lags internet data by two orders of magnitude. Shen is specific: one million hours of robot data is the threshold for balance with internet data.

"I believe the GPT-1 moment will come when embodied data itself can be scaled up, or at least reaches a similar order of magnitude to internet data. That's my idea of the GPT-1 moment." 00:09:07

"The robot data is about two orders of magnitude smaller than internet data." 00:08:55

"I think it should start at one million hours — when you can start saying the embodied data and the internet data are relatively balanced, where the model can learn both general knowledge and operation-related knowledge." 00:21:51

Causal (Unidirectional) Pre-Training Is Essential for Real-Time Robotics

One of Lingbo's hardest-won technical achievements is training a video generation model with purely unidirectional (causal) attention from scratch. Converting a bidirectional model to unidirectional destroys embedded knowledge. The lesson: architectural decisions for physical-world AI must be made at initialization, not retrofitted.

"When you forcibly convert it to unidirectional, the unidirectional part works but some knowledge is forgotten — the pre-training is wasted. The intuitive manifestation is: before, you could give it 100 sentences and each would generate a good video. Now maybe only 20-30% generate good videos; the rest it no longer understands, because that knowledge was destroyed." 00:58:05

"We made a decisive choice to abandon the bidirectional-to-unidirectional route, and designed it as unidirectional from the very start." 00:58:31

MoE Architecture for Robots: A Genuine Technical Capability Problem

Training a Mixture-of-Experts model for video generation that activates experts uniformly is described as a two-month, dozens-of-failures engineering problem — not a philosophical choice.

"Getting MoE trained properly is a capability problem, not an attitude problem." 00:55:57

"It took at least two months just to figure out MoE training... We estimate dozens of failures during those two months." 00:56:05

The Cross-Embodiment Brain Strategy: Software as the Constant, Hardware as Variable

Lingbo's strategic bet mirrors the Android/iOS model: one brain that runs on any body. The analogy to smartphones is made explicit — just as no single phone brand dominates, no single robot body will either.

"We hope that in the future, every robot will run Lingbo's model, or use Lingbo's brain. That was the fairly big decision we made early on." 00:16:52

"We hope our brain is cross-embodiment — universal. That's why in our training process we incorporated data from different embodiments. We want our brain to have experienced these embodiments during pre-training, to understand their characteristics, so it can quickly adapt when actually deployed in a specific embodiment." 00:27:24

The Brain Lags the Body — and That's the Current Opportunity

Shen makes the contrarian observation that in the current robotics landscape, the software/intelligence layer is actually behind hardware, inverting the conventional wisdom that China's strength is hardware and the US leads in software.

"From the current state of the industry, the brain is probably lagging behind the body right now." 00:01:38

"The current main contradiction should be: how do we first make robots capable of doing work properly." 00:25:14

Entrepreneurship in Deep Tech Requires Deliberate Bets Under Uncertainty

Shen frames startup strategy as "gambling" — not recklessness, but making asymmetric bets when the path is unclear, because if the path were clear, large companies with more resources would always win.

"Startups are definitely built on bets. No one knows which path will definitely succeed, and resources are always limited. So with limited resources and uncertain paths, you have to bet. Being decisive and daring to bet — I think this is a critical, critical point. Because if everyone knows which path leads to success, why would it be you who succeeds?" 00:33:09

2. Contrarian Perspectives

The Robot GPT-1 Moment Has NOT Happened Yet — And Most People Are Wrong About Why

The prevailing narrative is that robotics is in a rapid intelligence breakthrough. Shen argues the field hasn't even reached GPT-1 yet, and the reason is specifically data, not models or architectures.

"Even now I still feel we haven't reached that stage. Because there's still no good way to scale up data. Even with our 60,000 hours this time — honestly compared to last time, it's not an order-of-magnitude improvement. And these data are far from sufficient." 00:07:51

Simulation Data Is Overrated for General Robot Training

Against the industry trend of using simulation to cheaply generate massive training datasets, Shen argues that simulation data suffers from distributional homogeneity and can't yet handle the physical complexity of real manipulation tasks.

"Simulation data can instantly scale up to very large quantities in the things it's good at — that's indisputable. But from another angle, its distribution tends to be homogeneous. For example, simulating something like tearing open a plastic bag — I feel the physics engine still needs some time to iterate and upgrade." 00:35:25

Generalist Embodied AI Should Be Independent of Language Models — Not Built on Top of Them

Against the mainstream view that embodied AI is a downstream application of LLMs, Shen argues that the core problems of physical intelligence — how to do a task — don't require language understanding, and language models are merely a convenient entry point that will eventually be internalized.

"If you had 10,000 tasks and encoded them all as fixed codes, it might work — it just can't generalize in the task dimension. Language models solve the problem of demand generalization. But that's clearly not the core focus of embodied intelligence at this stage." 00:18:03

"Doing something well doesn't require re-understanding language. The step before it — translating language into a sequence of actions — may need to rely on language models. But I think embodied AI will continue to use large language models as a good entry point while growing its own model." 00:17:10

The Real Data Advantage in Robotics Belongs to China — and It's a Collection Problem, Not a Model Problem

Shen argues that the data advantage isn't about model training infrastructure but about the ability to physically collect embodied data at scale, where China's manufacturing and operational ecosystems give it a structural edge.

"Compared to the internet era, I think embodied data right now is in a window of opportunity — everyone is starting from the same place. China will definitely move faster than the US in data collection." 00:50:32

Tactile Sensors Should Be Redesigned Around AI Needs, Not Traditional Engineering Specs

The established industry assumption is that sensor quality metrics (precision, frequency) should be maximized. Shen argues AI models may need something entirely different — specifically, real-time response speed over precision.

"This thing — in the AI era — what the model needs most may not be high precision. It may need fast enough response. The next iteration of sensors probably needs to start from the model's requirements: which metrics to strengthen, which to weaken." 00:27:47

3. Companies Identified

Ant Lingbo (蚂蚁灵波)

Ant Group spinout focused exclusively on embodied AI foundation models (the "brain" layer), not hardware bodies. Released 6 open-source models in their second generation: Limbo Viren (geometric vision pre-training), Limbo Depth (spatial perception), VLA 2.0 (cross-embodiment action model), Video (MoE-native video generation), Word (causal pre-training), and VA 2.0 (the integrated robot-native foundation model). Pre-trained on 60,000 hours of real-robot and ego-centric data, up from 20,000 hours in v1. Supports 20+ robot embodiment configurations vs. 9 in v1, now including head, waist, and mobile base degrees of freedom. First Chinese lab to train a fully causal, MoE-native video-action model from scratch for physical-world use.

"We hope our brain is cross-embodiment — universal. That's why in our training process we incorporated data from different embodiments." 00:27:24

"We trained a technology designed purely for embodied control, with every parameter initialized completely randomly, using real-robot data, from scratch." 00:59:47

Physical Intelligence (π, PI)

San Francisco robotics AI lab. Described as the "OpenAI of robotics" by the host. Known for prioritizing zero-shot generalization — aiming for robots that need no fine-tuning for new tasks. Strong in real-robot teleoperation data.

"Pi seems to emphasize zero-shot more — they want to avoid fine-tuning, they want to be able to do everything." 00:13:43

"Physical Intelligence — they seem to be quite solid in real-robot teleoperation." 00:14:52

Generalist (Genie/Generist — likely referring to Physical Intelligence competitor or Google DeepMind Genie)

Referenced as having data pipelines significantly more mature than Lingbo's, with data possibly in the hundreds of thousands of hours or potentially millions of hours. Known for sophisticated teleoperation data scaling (likely referring to a leading US lab).

"Generist's data scale is much larger than ours. From publicly available information, it's probably in the hundreds of thousands of hours, and now I'm not sure if it's reached millions. Their data pipeline is more mature; they moved to pre-training earlier." 00:13:04

Google DeepMind

Cited for strong robotics research work influencing the field.

"Google DeepMind — their work on this side I think is also quite good." 00:49:40

ByteDance (字节跳动 / TikTok parent)

Shen's first employer after his PhD. Used his generative image editing research for TikTok special effects — age transformation, gender swap, style transfer. Described as a highly creative, effect-driven product culture.

"ByteDance was a good match for what I was doing — they were building TikTok, which is content-based. They needed a lot of special effects. I brought some interesting applications using work I did during my studies." 00:07:48

SenseTime (商汤)

AI unicorn where Shen interned during his undergraduate years at Tsinghua (2014 founding). Described as early in computer vision, not yet into generative AI during his time there.

"SenseTime was founded at the end of 2014, right at my junior year. It was right at Tsinghua's east gate. I went into AI-related fields and worked there until I graduated." 00:02:45

Unitree (宇树)

Chinese humanoid robot hardware company mentioned as having already achieved or near mass production.

"At least companies like Unitree and Zhiyuan have already gotten mass production running — the supply chain is also maturing as the industry develops." 00:19:05

Zhiyuan (智元)

Another Chinese humanoid robot company cited alongside Unitree as having reached mass-production-ready stage.

"At least companies like Unitree and Zhiyuan have already gotten mass production running." 00:19:05

Figure (飞格 / Figure AI)

US humanoid robotics company. Cited for demonstrating logistics applications in real-world deployment scenarios.

"Figure has gone and demonstrated logistics sorting — more and more companies are starting to talk about robot deployment, and they're not using traditional control methods anymore, they're using models." 00:20:49

4. People Identified

Shen Yujun (沈宇军)

Chief Scientist at Ant Lingbo. PhD from CUHK under Tang Xiaoou. Background in GAN-based image generation, image/video editing, and representation learning. Architect of Lingbo's embodied-native pre-training strategy, the causal MoE video generation model, and the geometric vision pre-training approach. First-time founder.

"I am very confident that embodied AI will definitely have its own model... How to scale up data well is a critically important point." 00:21:36

Tang Xiaoou (汤晓鸥)

Late founder of SenseTime and professor at CUHK (Chinese University of Hong Kong), Shen's PhD supervisor. Described as humorous and gifted at quickly identifying the core contradiction in complex, chaotic situations.

"Tang Xiaoou can quickly identify the main contradiction from a disordered, chaotic situation... During my thesis defense he asked me a question related to the industry — how should your work actually be used?" 00:44:38

5. Operating Insights

Diagnose Which Head of the "Two-Headed Snake" Is the Bottleneck Before Allocating Resources

Lingbo explicitly runs two parallel tracks — pushing the intelligence ceiling (research) and pulling from real deployment (product). The deployment track tells them which of the many possible model iteration directions to prioritize, preventing full-spectrum resource diffusion.

"The model is still early-stage — there may be ten or eight directions to iterate. But energy is limited; you can't do everything at once. Which capability to develop first is actually determined by the industry and by deployment." 00:18:59

Data Quality Pipelines Must Be Embodiment-Specific — Generic Cleaning Doesn't Work

A single universal data processing pipeline for cross-embodiment training is insufficient. Camera placement, joint configurations, and sensor setups differ so significantly between robot bodies that each embodiment requires tailored data handling. This is an operational moat.

"Many data processing methods are closely related to the specific embodiment. It's not that one data pipeline can completely handle every robot body... The camera placement may differ, the joint configurations may differ — these differences in machine configuration bring great difficulty to data processing." 00:38:08

Post-Training Data Requirements Dropped 5x Between Generations — Track This as a Leading Indicator of Model Maturity

The reduction in fine-tuning data needed per task (from ~100 demonstrations to ~20) and the ability to train multiple tasks with a single model rather than one model per task are quantifiable signals that pre-training quality is improving. These metrics are more actionable than benchmark scores.

"The previous version needed roughly 100 demonstrations per task. Now about 20 demonstrations is enough. Of course you still can't get to one or fewer — that probably requires the generalization capability to improve further." 00:07:06

Front-Load Sensor-Reality Into Training Data — Lab Cleanliness Is a Trap

Academic datasets are artificially clean; real deployment sensors are noisy. Training on clean data then deploying on noisy sensors causes systematic failure. The operational fix is to build training data pipelines that preserve and incorporate real sensor noise from day one.

"Sensor noise is something that absolutely cannot be ignored in real applications — it's naturally at odds with having clean data. So we committed to building from the sensor dimension from the start." 00:24:44

6. Overlooked Insights

The Semantic Encoder Alignment Trick: Baking Action and Semantic Signals Into the Visual Tokenizer

Shen briefly mentions that Lingbo's new video tokenizer (encoder) is trained with explicit semantic AND action alignment signals during the compression step — not just compression for reconstruction quality. This means the latent space is action-aware before the generative model even sees it. This is architecturally distinct from every standard video generation model (Wan, Sora-class models) and could be a meaningful moat if it proves out. It was mentioned in roughly two sentences and not highlighted by either speaker as a headline achievement.

"In order for the robot to better understand semantics and actions during learning, when training the visual compression model we added alignment of semantics and actions. In other words, during encoding the encoder can already perceive some semantic and action information. Having such an encoder, when later generating based on what it encoded, it becomes much better at understanding the connection between semantics and actions." 00:50:34

This is non-obvious because most world model research focuses on the generative backbone. If the tokenizer itself is action-conditioned, downstream policy learning could be substantially more sample-efficient — and this is exactly what Shen's "20 demonstrations for a new task" data point seems to confirm. No investor or researcher in the conversation flagged this as the structural reason behind the sample efficiency improvement.

Chinese Embodied Data Collection Has a Structural First-Mover Window That Is Closing

Shen notes that unlike internet data (where the US accumulated a massive head start), embodied robot data is at "zero" for everyone right now — it's a simultaneous starting line. China's manufacturing ecosystem, lower teleoperation labor costs, and hardware supply chain density mean it can collect embodied training data faster than any other geography. But this window only exists before one country's data flywheel compounds. This was mentioned in a single brief exchange and immediately moved past.

"Compared to the internet era, I think the embodied data — especially for embodied AI — is in a window of opportunity right now, where everyone is at the same starting point. China will definitely move faster than the US in data collection." 00:50:32

The investment implication is significant: any company building the infrastructure layer for embodied data collection in China (data collection protocols, annotation tools, data alliance platforms, teleoperation hardware) is competing for a window that may only be 18-36 months wide before network effects in data lock in a winner.