From Datasets to Deployment: Pan Pan Believes the Future of Robotics Depends on Structured Real-World Data

What motivated us was a very practical industry challenge. As embodied AI continues to evolve, many teams are making progress in model design and robotics hardware, but high-quality, scalable real-world data is still one of the biggest bottlenecks. Training embodied AI systems requires much more than static perception data. It requires data that reflects how intelligent systems perceive, decide, and act in real environments. 

We saw a clear gap there, and that was the starting point for developing our Embodied AI Data Solution. The TITAN award entry itself describes this challenge as the lack of scalable, real-world data for embodied intelligence development, which is exactly the problem we set out to address.

This direction is also closely aligned with Nexdata’s long-term goals. Since our founding in Singapore in 2011, Nexdata has focused on one core mission: helping organizations build advanced AI through better data. Over time, we have expanded from traditional AI data services into next-generation areas such as generative AI and embodied AI, because we believe the future of intelligent systems depends not only on better models, but also on stronger data infrastructure. 

Our embodied AI work is a natural extension of that strategy. It allows us to support customers in a field where data quality, realism, and scalability are becoming increasingly important.

What makes this especially meaningful for us is that we are not approaching embodied AI as a short-term trend. We see it as an important direction for the future of robotics and physical-world intelligence. Developing this solution was our way of contributing something practical to that future: not just ideas, but a real data foundation that can help embodied AI move closer to large-scale deployment.

Certainly. The core idea behind our award-winning entry is that embodied AI requires a fundamentally different data infrastructure from traditional AI. In conventional AI projects, the focus is often on recognition, classification, or language understanding. 

But in embodied AI, systems need to perceive, understand, decide, and act in the physical world. That means the data must capture not only what a robot sees, but also how it moves, manipulates objects, interacts with people, and completes tasks in real environments.

To address that need, we built our Embodied AI Data Solution as an end-to-end system rather than a single collection or annotation service. One of the most important foundations is our self-owned embodied AI data factory, which spans over 10,000 square meters and includes more than 200 humanoid robots and 300+ robotic hands and arms, covering mainstream robotic platforms. 

These include brands such as Unitree, Franka, Leju, and Linker, with support for a wide range of robot forms including humanoid robots, robotic arms, and dexterous hands. This gives us the ability to conduct large-scale embodied AI data collection under realistic and configurable physical-world conditions.

Another important aspect is our ability to support diverse real-world scenarios and task types. Our embodied AI environments cover settings such as supermarkets, pharmacies, factories, and auto repair shops, allowing us to collect data that is much closer to actual deployment needs. 

On top of that, we support a broad range of embodied AI tasks, including autonomous navigation, human-robot collaboration, object manipulation, long-horizon tasks, and force-feedback interaction. This is important because embodied AI performance depends heavily on whether the training data reflects the complexity of real tasks, not just isolated lab actions.

What also makes our solution unique is its multimodal and model-oriented design. We support different embodied AI data formats such as ego-exo action video annotation, motion-related data, sim-to-real data support, 3D-related data, voice interaction, and tactile-force data. We also provide specialized annotation services for embodied AI tasks, including pose annotation and VLA-related data processing. 

In other words, we are not just collecting raw data — we are building structured, training-ready data pipelines that can better support the development of next-generation robotics and embodied intelligence systems.

If I had to summarize the innovation in one sentence, I would say this: our solution turns embodied AI data production from a fragmented manual process into a scalable, structured, and real-world-oriented infrastructure system. That is what makes it valuable for robotics companies and AI teams that want to move from experimentation to practical deployment.

My contribution was less about one technical decision and more about helping shape the project in the right direction from the beginning. In embodied AI, it is easy to focus only on the hardware or the model side, but we believed early on that data infrastructure would be one of the key factors determining whether embodied AI could scale in real-world applications. 

That judgment shaped how we approached the project and where we invested our resources. The TITAN-winning entry itself reflects that direction, as it focuses on solving the shortage of scalable, real-world data for embodied intelligence.

From a leadership perspective, an important part was aligning multiple capabilities into one integrated solution. Embodied AI data projects involve not only collection, but also environment design, task planning, annotation standards, quality control, and delivery workflows. At Nexdata, we already had experience in integrated AI data services, including ready-to-use datasets, data collection, labeling customization, and platform-based delivery, which gave us a strong foundation for building this solution in a more systematic way.

Another key contribution was making sure the project was built with execution quality in mind. For a solution like this, success depends on whether it can operate consistently at scale, not just whether it sounds innovative. Nexdata’s quality system, including ISO 9001-certified quality management and multi-round quality assurance, helped us bring that discipline into the project. That made it possible to translate an ambitious idea into a practical, repeatable service capability.

So overall, I would say my role was to help connect vision with execution: identifying the right opportunity, organizing the right teams and capabilities around it, and ensuring that the final solution was both innovative and usable for real embodied AI development.

Our innovation addresses a fundamental challenge in embodied AI: the lack of scalable, real-world, and training-ready data for systems that need to perceive, decide, and act in physical environments. As our TITAN-winning entry explains, traditional data approaches are often built around static perception tasks or synthetic scenarios, which are not enough for training embodied AI systems that must operate reliably in real-world settings.

What we are solving is not just a shortage of data volume, but a gap in the entire data production process. In many existing workflows, embodied AI data collection is fragmented, task-specific, and difficult to scale. Data may come from limited lab settings, isolated robot configurations, or disconnected manual workflows, which makes it hard to support model training efficiently. Nexdata improves this by building an integrated data service framework that combines off-the-shelf datasets, real-world data collection, and high-quality data annotation for embodied intelligence model training.

In practical terms, our solution improves existing processes in three ways. First, it makes data collection more realistic by using configurable real-world environments such as supermarkets, pharmacies, factories, and auto repair shops, rather than relying only on simulation or narrow lab conditions. 

Second, it makes data production more scalable through a closed-loop workflow that covers collection, cleaning, synchronization, annotation, quality control, and iterative optimization. 

Third, it makes the output more useful for downstream model development by structuring data for perception, decision-making, and control tasks, including support for imitation learning, reinforcement learning, and Vision-Language-Action models.

So overall, I would say our innovation improves the process by turning embodied AI data work from a fragmented manual effort into a more standardized, scalable, and model-oriented system. That is what helps reduce the gap between experimentation and real-world deployment.

I believe what makes our innovation stand out is that it is not a single-point service, but a full embodied AI data infrastructure solution designed for real-world robotics development. The goal of our TITAN-winning entry is to solve the lack of scalable, real-world data for embodied intelligence, especially for systems that need to combine perception, decision-making, and physical control. That already sets it apart from more traditional AI data workflows that are often centered on static perception or synthetic scenarios.

One of the most important differentiators is our dedicated embodied AI capability. Nexdata provides tailored services for embodied AI applications across visual perception, motion control, and interaction with the environment, rather than treating robotics data as an extension of conventional annotation work. Our service scope includes ego-exo action video annotation, motion-related data, sim-to-real support, 3D-related data, voice interaction, and tactile-force data, which reflects the multimodal nature of embodied intelligence.

Another key strength is our real-world collection environment. Nexdata has built an embodied intelligence data factory with configurable physical settings such as supermarkets, pharmacies, factories, and auto repair shops. This is important because embodied AI systems need to learn from realistic operational contexts, not just controlled lab demonstrations. That kind of environment support makes the data more valuable for downstream training and deployment.

What also stands out is that we focus on structured, model-oriented output, not just raw data generation. Our embodied AI solution is built to support downstream model development more directly, including applications related to imitation learning, reinforcement learning, and Vision-Language-Action-style workflows. In other words, we do not just help customers collect data — we help them build data pipelines that are more aligned with how embodied AI systems are actually trained and evaluated.

Finally, I would say our innovation stands out because it reflects Nexdata’s broader strength as a long-term AI data partner. Nexdata is a Singapore-founded AI data solutions company with experience across ready-to-use datasets, custom collection, annotation, and curation services, and that broader foundation allows us to bring industrial-scale thinking into embodied AI, which is still an emerging field for many companies.

Our company and team played a central role in turning this idea into a practical and scalable solution. Embodied AI data is not something that can be delivered by a single function alone. It requires the coordination of infrastructure, data operations, project management, annotation expertise, and quality control. At Nexdata, we already had a strong foundation in integrated AI data services, including off-the-shelf datasets, real-world data collection, and high-quality annotation, and that allowed us to build this solution in a much more systematic way.

What made the difference was the combination of long-term company investment and cross-functional execution. Nexdata built dedicated embodied AI capabilities, including a specialized embodied AI data service framework and real-world scenario support for tasks related to perception, decision-making, and control. That means our team was not simply responding to one project requirement — we were building a full capability set around embodied intelligence.

Our team’s contribution was especially important in connecting the different stages of delivery. From environment setup and data collection design to annotation workflow, resource allocation, and quality assurance, each part had to work together. Nexdata’s broader organizational strengths — including experienced data scientists, project managers, linguists, resource development teams, and established quality systems — helped make that possible.

So overall, I would say the company’s role was to provide the platform, infrastructure, and long-term strategic commitment, while the team’s role was to translate that foundation into an executable, high-quality embodied AI data solution that customers can actually use.

One of the biggest challenges was that embodied AI data is much more complex than traditional AI data. It is not just about labeling what appears in an image or audio clip. We had to think about how to capture and organize data related to perception, movement, interaction, task execution, and real-world environments in a way that would actually be useful for model training. That made the development process more demanding in terms of workflow design, annotation standards, and quality control.

Another challenge was balancing realism with scalability. In embodied AI, real-world data is extremely valuable, but real-world collection is also harder to standardize. Different environments, tasks, devices, and interaction patterns can introduce a lot of variability. So one important part of the work was making sure that our solution could support realistic scenarios while still maintaining consistency and delivery efficiency.

What helped me most was staying structured and execution-focused throughout the process. I tend to break complex problems into clear operational steps, align the right teams around each stage, and keep the project centered on what will create real value for customers. In a project like this, that means not getting lost in technical complexity, but constantly asking whether the solution is scalable, usable, and aligned with actual embodied AI development needs.

I also believe communication played an important role. Because embodied AI projects involve multiple functions, from collection and annotation to platform support and project delivery, it was important to keep everyone aligned around the same objective. My role was to help connect strategy with execution, make decisions when priorities needed to be clarified, and ensure that the final solution was both innovative and practical.

We hope our innovation can help accelerate embodied AI from early experimentation to more practical and scalable real-world deployment. Today, one of the biggest constraints in embodied AI is not only model capability, but also the availability of high-quality, structured, and real-world data. Our TITAN-winning entry was built around that challenge — to provide scalable data support for systems that need to combine perception, decision-making, and physical control in real environments. 

For the industry, we hope this solution can help strengthen the data foundation behind next-generation robotics and embodied intelligence. Embodied AI is moving beyond isolated research settings, and that means the industry needs more standardized, scalable, and model-oriented data pipelines. If our work helps companies reduce the gap between lab development and real-world application, then we believe it will contribute real value to the future of robotics. 

For customers and AI teams, our goal is to make embodied AI development more efficient and more practical. By combining real-world data collection, structured annotation, and support for different embodied AI scenarios and task types, we hope to help teams spend less time dealing with fragmented data challenges and more time improving model performance and deployment readiness. 

Ultimately, we hope our innovation shows that data infrastructure is not a secondary part of embodied AI — it is one of the foundations that will shape how far the industry can go. That is the impact we would like to have: helping make embodied AI more trainable, more scalable, and more applicable in the real world.

Winning this award reflects a belief we have held for a long time: technological progress is not only about making models or hardware more advanced, but also about building the infrastructure that allows those technologies to work in the real world. In embodied AI especially, progress depends on whether systems can learn from scalable, real-world data for perception, decision-making, and physical control. That is exactly the problem our award-winning solution was designed to address.

For us, innovation is most meaningful when it is practical, scalable, and usable. Nexdata has focused on AI training data since 2011, and our broader vision has always been to help organizations build stronger AI systems through better data foundations rather than treating data as a secondary step. This award is meaningful because it recognizes that data infrastructure itself is a real driver of innovation, especially in emerging areas like robotics and embodied intelligence.

It also reinforces our view that the future of innovation will be increasingly interdisciplinary. In embodied AI, breakthroughs do not come from one layer alone. They come from the combination of hardware, models, environments, and high-quality data systems.  This award affirms our belief that real innovation is not just about what AI can do in theory, but about what it can reliably do in the real world.

One of the main challenges during the development process was the complexity of embodied AI data itself. Unlike traditional AI data projects, embodied AI requires us to capture not only perception data, but also movement, interaction, task context, and environment-related information in a structured way. This means the data pipeline has to support much more than collection alone — it also needs to handle synchronization, annotation standards, quality control, and downstream usability for model training.

Another challenge was building a solution that was both realistic and scalable. Real-world embodied AI data is valuable because it reflects actual physical environments, but that also makes it more difficult to standardize. Different robots, tasks, environments, and interaction patterns can easily create inconsistency if the workflow is not designed carefully.

We addressed these challenges by taking a systematic approach. First, we built dedicated collection environments and task workflows to improve consistency at the source. Second, we combined data collection with structured annotation, processing, and quality management so the output would be more useful for training and evaluation. Third, we relied on close collaboration across teams, because a project like this requires infrastructure, operations, annotation expertise, and delivery capabilities to work together.

In short, we overcame these challenges by focusing on process design, operational discipline, and cross-team execution. That was what allowed us to turn a complex embodied AI requirement into a practical and scalable data solution.

I believe our innovation can help push the embodied AI industry toward a more mature and scalable stage. Today, many companies are making progress in robotics hardware and model development, but data infrastructure is still one of the biggest constraints. 

Our solution is designed to address that gap by providing more structured, scalable, and real-world-oriented data support for embodied intelligence. That is also how our TITAN-winning entry is positioned: as a response to the lack of scalable real-world data for perception, decision-making, and control in embodied AI.

In the future, I think the industry will rely much more on standardized and model-oriented data pipelines, especially as embodied AI moves from research environments into commercial applications. Our role is to help make that transition easier by giving robotics and AI teams access to data services that are closer to real deployment needs. If we can help reduce the gap between experimentation and real-world use, then I believe our innovation will have a meaningful impact on the future of the industry.

What excites me most right now is the convergence of multimodal AI, Vision-Language-Action models, and embodied intelligence. We are seeing AI move beyond understanding text, images, or speech separately, toward systems that can connect perception, reasoning, and action in the physical world. 

That trend directly influences our work, because it raises the requirements for training data. Customers no longer need only isolated perception datasets; they increasingly need multimodal, task-oriented, and interaction-centered data that can support embodied AI development more effectively.

I am also very interested in areas such as sim-to-real transfer, human-robot collaboration, and tactile or force-related data, because these are all becoming more important as embodied AI systems are expected to perform in more realistic environments. Nexdata’s embodied AI services already cover areas such as ego-exo action video annotation, motion-related data, sim-to-real support, 3D-related data, voice interaction, and tactile-force data, so these technology trends are very closely connected to the direction of our work.

My advice would be: focus on the real bottleneck, not just the most visible part of the problem. In emerging industries, people are often drawn to the most exciting layer, whether that is the model, the interface, or the hardware. But transformative progress often depends on solving the less visible infrastructure problems underneath. In our case, embodied AI cannot scale without better real-world data systems, so that is where we chose to invest.

I would also say that transformative ideas need both vision and discipline. It is important to think long term, but it is equally important to build something that is practical, repeatable, and useful. A good idea becomes valuable only when a team can execute it consistently and translate it into real outcomes. For teams working on ambitious innovations, that balance between imagination and execution is often what makes the biggest difference.