Accelerating Embodied Intelligence with Multimodal Data

Insights from the 2025 01VC Forum on Smart Manufacturing and Cross-Border Expansion

Embodied intelligence is one of the most significant technology revolutions underway—poised to redefine how humans interact with the physical world. But the path to deployment faces a major bottleneck: the lack of high-quality, large-scale robot-environment interaction data.

At the 2025 01VC Forum on Smart Manufacturing and Cross-Border Expansion, the roundtable centered on multimodal data for embodied AI.

01VC Partner Nick Zou moderated a conversation with three portfolio company founders:

• Fan Yu, Founder of Transfer Technology
• Guan Saixin, Founder of ZingTech
• Tian Ye, Founder of RoboScience

The discussion explored core technical progress, data challenges, multimodal fusion, and commercialization pathways.

Part 1: Building the Eyes, Skin, and Brain of Embodied AI

Nick Zou, 01VC:

Embodied intelligence is one of the hottest areas in tech. Today we’re joined by three founders who are building its foundational components: Fan Yu from Transfer Technology(robotic vision), Guan Saixin from ZingTech (robotic task sensing), and Tian Ye from RoboScience (robotic cognition). Could each of you share an overview of your company and where you’re focused?

Fan Yu, Transfer Technology:

Transfer Technology develops 3D vision systems for industrial and robotics applications. We used to describe our system as a robot’s “eyes and brain,” but today we call it the “eyes and cerebellum.” We recently launched the world’s first smart camera with built-in compute, eliminating the need for external processors. This innovation has reduced system costs by 70–80% and caused a real stir in the industry. We're also building out VLM and VLA models to enable deeper embodied perception.

Guan Saixin, ZingTech:

We focus on tactile (task sensing) technology. Task sensing data is increasingly in demand—second only to vision in information richness. We're working across robotics, industrial machinery, and automotive safety. For example, we’re collaborating with automotive manufacturers to develop tactile-enabled safety systems. Electronic skin has broad applications, well beyond humanoid robots.

Tian Ye, RoboScience:

Our work is centered on building a robot's “brain.” Traditional robots must be explicitly programmed to complete each task. Our goal is to develop a general-purpose robotics model that can control any robot, in any environment, to manipulate any object—without manual reprogramming. This would transform robots from tools into intelligent collaborators that amplify human productivity.

Part 2: Data Bottlenecks and Simulation Challenges

Nick Zou, 01VC:

Data is widely viewed as the biggest bottleneck for embodied intelligence. Unlike the internet, there’s little high-quality robot-interaction data available, and the scale is orders of magnitude below what’s needed. How can we solve this?

Fan Yu, Transfer Technology:

VLA (Vision-Language-Action) models mirror human cognition, which can be broadly divided into two systems. For example, when driving, route planning from A to B represents a ‘slow thinking’ process—deliberating between highways or city roads. In contrast, the rapid responses required while operating a vehicle fall under ‘fast thinking.’

In the realm of embodied intelligence and robotics, a similar dual-model structure exists. The ‘slow thinking’ component—task decomposition through language or semi-linguistic models—is relatively mature from a technological standpoint. The ‘fast thinking’ component—requiring real-time data feedback and closed-loop visual control—remains significantly more complex. Current advances in robotics and embodied learning are focused precisely on these high-frequency, real-time interaction data.

To address the challenge of limited embodied data, we can draw parallels from the evolution of large language models. Early-stage LLMs required millions of data points for effective training. Today, through refined data curation and reinforcement learning techniques, models can achieve strong performance with just tens of thousands of samples. Applying similar principles—particularly reinforcement learning—could greatly reduce the data requirements for embodied intelligence systems.

Potential approaches include:

• Developing simulated environments for embodied training, enabling models to acquire foundational capabilities in a cost-effective and scalable manner.
• Utilizing 3D vision technologies to structure real-world environments—identifying key elements such as chairs, people, or cars—and transforming them into virtual representations, allowing direct deployment of models trained in simulation.

This strategy is inspired by the training of quadruped robots, where simulation-trained models can be effectively transferred to the real world. However, unlike quadrupeds, which do not require precise manipulation, embodied intelligence systems often demand coordinated control across hand-eye-foot systems with high accuracy—rendering prior approaches insufficient. By combining simulation, 3D structuring, and reinforcement learning, we may be able to overcome the current bottlenecks in embodied data acquisition and accelerate real-world deployment.

Tian Ye, RoboScience:

We share Mr. Fan’s perspective that virtual environments are currently the most scalable and efficient approach to collecting embodied intelligence data. Traditional methods—such as teleoperated robot data collection—yield just 200–300 samples per day, making them fundamentally inadequate for training general-purpose embodied AI systems at scale.

The industry’s core challenge lies in the “sim-to-real gap”—models trained in simulation often suffer a sharp drop in performance when deployed in the real world. For quadruped robots, this gap is manageable due to predictable morphology and well-defined physics. But for general-purpose robots performing complex manipulation tasks—interacting with diverse, unknown objects like cups, utensils, or clothing—precise simulation becomes exponentially harder.

Most robotics companies rely on generic, open-source, or commercial physics engines that are not built for robotics. As a result, the synthetic data they produce lacks the fidelity needed for effective model training, especially at scale.

To address this, we’ve developed our own proprietary simulation engine. Unlike off-the-shelf options, our engine can model nuanced physical interactions—accurately simulating forces, friction, elasticity, and more—based on real-world physics. This dramatically enhances the realism and transferability of training data, which we view as the first critical step in narrowing the sim-to-real gap.

We pair this with a two-phase training strategy:

• Pre-train in high-fidelity simulation, leveraging large-scale, structured virtual data
• Fine-tune with minimal real-world data to bridge deployment gaps cost-effectively

Our system architecture includes both a planning layer and an execution layer. The planning layer benefits from broad internet-scale video data (e.g., demonstrations of human and robotic actions). The execution layer—responsible for real-world action—is where the gap is most pronounced and where our proprietary simulation capability provides the strongest edge.

In our view, the ability to control both the algorithm and the simulation infrastructure is a key differentiator in embodied AI. It not only enables faster iteration and higher-quality data, but also lays the groundwork for scalable deployment across real-world applications—from home robotics to industrial automation.

Nick Zou, 01VC:

Most models today use vision and language inputs (VL), but for embodied intelligence, VL isn’t enough. We’re now seeing VLTA—where T stands for Task. Guan, what value does tactile data bring?

Guan Saixin, ZingTech:

We’ve recently been co-developing products with several industrial clients and observed a notable gap: over 80% of industrial robotic arms and grippers currently lack any form of tactile sensing. This is a real and widespread limitation. In comparison, developing tactile systems for humanoid robots is even more challenging.

This situation stems from a broader industry reality—while vision and language models are progressing, tactile sensing is often treated as an afterthought. However, we’re now starting to see growing demand for tactile capabilities in adjacent sectors. For example, some of our industrial clients are actively adding tactile sensors to their grippers, and these use cases provide transferable technologies and experience for the embodied AI and humanoid robotics space. In other words, there is real horizontal applicability of tactile innovations, making these industrial applications a high-leverage entry point.

When it comes to bridging simulation and real-world deployment, tactile-specific challenges—such as grip force and sliding friction—present further complexity. Today, there are no widely accepted industry standards for tactile sensors in these applications. That’s why we’re focused on foundational work: building core tactile sensor technologies first, akin to “making bricks before building walls.” The advantage of tactile sensors is their long-term scalability—once developed, they can serve a wide range of use cases beyond their initial application.

We recently came across a humanoid robotics team that set human-level tactile performance as a baseline requirement for their robots. While the ambition is commendable, the economic reality was overlooked—building to that spec today would render the solution commercially unviable. As a company focused on scalable, large-area tactile solutions, we differentiate from mainstream by targeting high coverage at low cost. Our goal is to keep the cost of tactile skin at around 10% of the total cost of a dexterous hand or robotic arm, which we believe is the right benchmark for mass adoption.

From a cost structure perspective, this is especially critical if humanoid robots are to be deployed in industrial or household settings. In factories, the typical ROI timeframe for automation is 1.5 years. With monthly labour costs ranging from RMB 5,000 to 7,000, that translates to an acceptable robot price point of ~ RMB 150,000. For home service robots, a target price of RMB 100,000 would require even more aggressive cost optimization. Ideally, the cost of a robotic hand would be around RMB 10,000, with tactile skin accounting for ~RMB 1,000.

To reach this cost-performance threshold, we believe the key is to first drive down sensor costs in adjacent sectors, and then apply those gains to humanoid robotics. This is similar to how smartphone companies entered the automotive industry—leveraging existing supply chains and accumulated know-how to make vehicles smarter and more cost-effective.

For us, humanoid robots remain the core long-term focus. But before the market reaches scale, we must first solve for tactile sensor cost and scalability. This strategy is aligned with what Mr. Fan is doing—starting with industrial robots and transitioning to humanoids—which we believe is a pragmatic and capital-efficient path forward.

Part 3: Commercialization & Market Readiness

Nick Zou, 01VC:

The embodied intelligence sector has seen a surge of activity over the past two years, with hundreds of startups entering the space. Despite the momentum, the industry remains in the early stages of commercialization, with most companies still searching for product-market fit (PMF). From an investor’s perspective, a critical question now is: how far are we from seeing clear PMF emerge in this category—and which segments are likely to cross that threshold first?

Fan Yu, Transfer Technology:

Cost-efficiency is a critical consideration for industrial adoption. The clients we serve in manufacturing environments rigorously assess return on investment (ROI) for each automation scenario. Before committing to purchases—whether for cameras, robotic arms, or other automation equipment—buyers typically calculate payback periods. If labor savings can offset the upfront investment within two years, they are generally willing to proceed. As a result, price sensitivity remains high across the board.

Embodied intelligence products, such as dual-arm systems or mobile platforms, often carry higher costs. This raises a key question: under what conditions will factories be willing to adopt these more advanced systems?

Looking ahead, if general-purpose humanoid robots were to become mainstream, their hardware architecture could potentially align closely with the automotive supply chain—enabling cost reduction through economies of scale and driving generalization via software. However, two major challenges remain:

• Can the industry converge on a standardized hardware configuration for humanoid platforms?
• Even if standardization is achieved, how can we systematically collect real-world data and train high-performance foundation models throughout deployment?

These are critical execution bottlenecks that will determine whether embodied AI systems can scale beyond R&D into commercially viable platforms.

Guan Saixin, ZingTech:

Over the past few years, our company has built a broad portfolio across service robotics—including robotic vacuum cleaners, robotic lawnmowers, pool cleaning robots and various humanoid robot form factors. While recent investor attention has largely focused on humanoid robots, the reality on the ground is that the broader industry is not yet fully prepared for scaled deployment.

In this context, we believe a pragmatic and sustainable strategy is to maintain long-term focus on humanoid robotics, while ensuring near-term survivability and self-sufficiency. The key challenge—for us and for the industry at large—is identifying a viable development pathway and monetization model before humanoid robots achieve true commercial maturity.

It’s also important to recognize that the hype cycle around humanoid robotics is volatile. Just two years ago, market interest was minimal, with few believing in its potential. Last year and the first half of this year, sentiment swung dramatically in the other direction. Our belief is that to succeed in this field, companies must commit deeply to the long-term vision—while maintaining strategic discipline: avoiding irrational exuberance during upcycles, and staying resilient during downturns.

History offers no shortage of examples where waves of companies surged and faded in emerging technology sectors. But we remain grounded in the conviction that if you can clearly envision the inflection point—be it in 3, 5, or even 20 years—you must ensure your company can survive and continue building until that moment arrives. That means staying close to leading customers, co-developing real products, and maintaining the capability to execute when the market eventually accelerates.

Tian Ye, RoboScience:

The long-term vision for embodied intelligence is to enable general-purpose capabilities—robots that can perform a wide range of tasks. While the value potential is significant, progress must be incremental and grounded in real-world use cases.

Achieving this requires aligning current technical capabilities with deployable scenarios—defining the data needs of each task, training fit-for-purpose models, and integrating them into robotic systems.

To scale efficiently, models should be designed for adaptability—for instance, supporting multiple robot types broadens deployment opportunities and accelerates data collection.

This is ultimately a strategic balance: enabling practical deployments today while building the software, data, and hardware foundations needed for long-term scalability.

Nick Zou, 01VC:

Thank you Fan, Guan, and Tian for the incredible insights. We look forward to future collaborations across your companies—and to accelerating the embodied intelligence industry together.

Closing Summary

As embodied intelligence moves from concept to capability, the companies shaping its foundation — ZingTech, Transfer Technology, and RoboScience — are tackling the hardest technical and commercial challenges head-on: multimodal data collection, simulation fidelity, sensor standardization, and scalable product paths. While the promise of general-purpose humanoid robots remains long-term, the path forward is being paved today through targeted applications, capital-efficient innovation, and cross-industry technology transfer.

01VC remains committed to supporting this next generation of deep tech founders—those who aren’t just betting on the future of robots, but building it brick by brick. The embodied revolution is just beginning.

‍