2026-07-29 18:06

Introduction: World Labs bets on the Real-to-Sim-to-Real closed loop, eschewing hardware development in favor of synthetic data to bridge the gap in embodied intelligence training and evaluation

On July 22, World Labs announced the acquisition of SceniX, a top-tier robotics simulation startup. The merger of two elite R&D teams marks a pivotal shift in the competition for embodied intelligence—from “pure language models” to “physical world models”

Immediately following the announcement, a16z partner Martin Casado engaged in a deep dialogue with Li Feifei, co-founder and CEO of World Labs, and Li Yunchu, co-founder of SceniX. The conversation centered on the full Real-to-Sim-to-Real training loop, systematically unpacking cutting-edge topics including world models, robot foundation models, synthetic data, and evaluation frameworks.

Li Feifei and Li Yunchu emphasized that the future of AI must go beyond language and text understanding lacking physical constraints—it must endow machines with “spatial intelligence”, enabling perception, reasoning, and real-time interaction with physical space. Training embodied intelligence demands not only a fundamentally different underlying logic from LLMs but also a departure from the current industry’s blind obsession with pure video prediction models and general-purpose humanoid robots.

Key insights:

1. Spatial intelligence is AI’s next frontier, with robotics as its optimal physical embodiment. AI evolution will inevitably transcend mere text interaction, advancing toward world models capable of perception, reasoning, and spatial interaction. While virtual world creation remains vital, taking action in the real physical world constitutes the core application scenario for spatial intelligence.

2. Unlocking the robotics Scaling Law hinges on the Real-to-Sim-to-Real data closed loop. Unlike large language models, which can leverage massive internet-sourced data, robotics cannot directly access abundant physical data. High-fidelity digital modeling and simulation enable the mapping of real environments into digital twins, generating infinite synthetic data—this is the key to overcoming data scarcity and enabling efficient evaluation.

3. Pure video models lack physical and spatial consistency, rendering them ineffective as robotic feedback signals. Many popular video models produce outputs inconsistent across space and time, even exhibiting hallucinations like objects vanishing mid-scene. Without genuine physical constraints and interaction feedback, robots cannot learn valid strategies.

4. Robot foundation models will evolve into forward simulators and policy models outputting actions. Future robot base models will abandon traditional text-based interfaces, evolving instead into forward simulators predicting environmental changes and policy models directly generating concrete physical actions in response to target instructions.

5. Simulation plays an irreplaceable role in counterfactual reasoning and extreme boundary testing. While real-world data is essential, simulation enables cost-free rehearsal of high-risk, extreme, or rare scenarios in digital space—significantly accelerating training while enhancing system reliability.

6. Avoid blind hardware fabrication; prioritize building model- and embodiment-agnostic infrastructure. Success in the embodied intelligence ecosystem lies in specialization. A universal simulation and evaluation platform that is agnostic to underlying algorithms and hardware embodiments (single-arm, dual-arm, dexterous hands) will accelerate the evolution of all robot "brains."

7. Rather than chasing general-purpose humanoids, tackling semi-structured environments—the "dirty, hard work"—is the rational path. Robotics deployment will follow a natural progression: from factory assembly lines to warehouses and restaurants, eventually reaching homes. Solving specific pain points with customized hardware is far more feasible commercially and technically than forcing expensive humanoid forms onto every use case.

Transcript of the Conversation:

1. Pioneering a New Frontier: From Spatial Intelligence to Acquiring a Robotics Startup

Host: For listeners unfamiliar with the background, could you briefly introduce what World Labs does?

Li Feifei: Sure. World Labs is a frontier model lab founded two years ago. We are building the next frontier of AI—what we call spatial intelligence. Spatial intelligence refers to AI's ability to generate, understand, reason, and interact with physical or virtual spaces. A key pathway to achieving this is constructing large-scale world models—a core focus of World Labs.

Host: You mentioned early on the capabilities of machine perception, reasoning, and acting in space. I used to think "acting in space" was distant, yet now you’ve acquired a robotics company. Can you discuss the timing and motivation behind this acquisition?

Li Feifei: First, acting in space doesn’t necessarily require robots. In creative domains like visual effects, gaming, or design, many scenarios allow for action within virtual spaces. Our core belief at World Labs has always been that our lived reality can be multidimensional—we build technologies enabling developers to act across diverse spaces.

That said, acting in the physical world is one of the most exciting and critical capabilities in the future AI landscape, and robotics is a key carrier of this capability. World Labs has long considered robotics a crucial application for spatial intelligence and world models. Inviting the SceniX team into World Labs is part of fulfilling our long-term vision and mission—a direction we’ve consistently pursued.

2. Origins of Collaboration: SceniX Team Background and the Serendipitous Partnership

Host: Li Yunchu, as co-founder of SceniX, could you briefly share your background and what SceniX does?

Li Yunchu: Of course. I’m Li Yunchu, co-founder of SceniX, and also an assistant professor at Columbia University.

Li Yunchu: My research began during my PhD at MIT, followed by postdoctoral work at Stanford under Li Feifei.

Host: The world truly is small.

Li Yunchu: Indeed. Throughout my academic and professional journey, my goal has been singular: helping robots better perceive and interact with the physical world. I’m a pragmatist—I want robots I build to function reliably in real environments.

For SceniX, we identified a major bottleneck in general robotics development, particularly in training and evaluation. Thus, we developed a pipeline called Real-to-Sim-to-Real—a complete technical loop that maps real environments precisely into the digital world. By “precise,” we mean any change in the digital world synchronizes with the real one. This allows us to replace all real-world data and evaluation needs with scalable synthetic data generated in the digital twin. That’s how SceniX was born. We’ve gathered top talent from robotics, robot learning, simulation, and rendering to build this Real-to-Sim-to-Real tech stack, addressing a critical industry bottleneck.

Host: It’s wonderful that you two are reuniting.

Li Feifei: Yes, there’s a fascinating story. You might assume we’d collaborated before, so we’d have discussed integration extensively in private—but the opposite is true. They initially approached World Labs as customers. Last winter, around November or December, we launched the first version of Marble, our generative foundational model. SceniX immediately registered as a customer. I didn’t even know whose account it was. When I called Li Yunchu, I was surprised to discover it was his startup project—and we realized the immense synergy between our teams.

3. Breaking the Data Impasse: The Power of Generative 3D and Simulation Synergy

Host: Could you briefly explain what Marble is?

Li Feifei: Marble is the codename for the foundational model World Labs has been training and iterating on. The publicly released version of Marble can take multimodal inputs—single images, multiple images, or text—and convert them into geometrically consistent 3D worlds, representable via Gaussian Splatting or 3D meshes. What the SceniX team is doing addresses a highly challenging problem in robotics: data scarcity. Both training and evaluation data are severely limited—unlike LLMs, where vast amounts of data are readily available online. We know unlocking scaling laws is essential for breakthroughs in robotics. But where does the data come from? This is a profound challenge every robotics R&D team wrestles with.

Host: The synergy between your teams is indeed fascinating. Both companies boast exceptional technical teams—how much overlap exists in your business, and is this integration primarily about expansion? Could you elaborate?

Li Feifei: Our businesses are highly complementary and share a common mission. Li Yunchu is one of three technical co-founders; the other two are Professor Changi Zhang from Columbia University, a leading expert in simulation with experience at Weta Digital and Tencent, bringing visual effects expertise and entrepreneurial experience; and Sunonni, an outstanding engineering lead who previously worked at a startup acquired by Amazon, amassing deep computer vision technology stack experience. When we began deep discussions, I realized SceniX offered perfect talent supplementation for World Labs: first, they possess full-stack robotics capabilities—from modeling to hardware. During his postdoc at Stanford, Li Yunchu already held a faculty position, though I hoped he’d stay longer. He’s a full-stack robotics researcher, and the student team he and Changi built at SceniX is exactly the talent pool World Labs previously lacked. Second, Changi’s team has extraordinary simulation capabilities—Changi is a senior simulation researcher and expert, and World Labs’ work heavily relies on integrating into simulation environments.

Li Feifei: What SceniX lacks naturally are generative models and strengths in computer vision and 3D reconstruction—precisely World Labs’ core competencies, which SceniX desperately needs. Together, these capabilities create a much more complete technical loop.

Host: Your motivation for expanding into robotics is clear, Li Feifei. Li Yunchu, stepping away from founding and joining World Labs is a major decision. How did you assess alignment with World Labs, and why did you make this choice?

Li Yunchu: Initially, we planned to continue independently. But after deep conversations with Li Feifei, we saw the enormous synergistic potential—our teams together are a perfect match. From SceniX’s perspective, our Real-to-Sim-to-Real approach fundamentally involves reconstructing environments—capturing appearance, geometry, and dynamic changes (how environments respond to actions). Currently, tensor-based reconstruction remains costly. World Labs has deep expertise in sparse reconstruction and generative capabilities, offering huge opportunities to leverage Marble and other World Labs technologies for efficient environment reconstruction and modeling.

4. The Future Foundation for Robots: Multimodal “Omni-Models” and Action Prediction

Host: Can we expect World Labs to launch a general-purpose foundation model for robotics?

Li Feifei: World Labs is indeed building foundational models. You know, we’ve been developing base models. As technology evolves, the most exciting base models today are omni-models—capable of handling multimodal inputs and producing multimodal outputs. What is a robot’s foundation model? It will likely incorporate actions as outputs. Beyond world state prediction, it may well output actions—this possibility is absolutely not ruled out.

Li Yunchu: Exactly. A robot’s foundation model must inherently be multimodal, processing video frames, text, images, depth, and more—with actions being a critical component. If current frame and action are inputs, it functions as a forward simulator predicting how the environment changes upon applying a given action; when actions are outputs, it becomes a policy model predicting what actions a robot should take in the real world to achieve a given goal. Such omni-models bring tremendous value to robotics—they help understand how to model environments and solve how to act within them. They can serve as a base for fine-tuning toward specific robotic applications, ensuring the reliability and efficiency clients expect.

Host: As a non-expert investor, I’ll ask: I see many robotics companies now using video models directly. This differs sharply from your 3D and simulation path. How do you contrast this popular video-only approach with your vision?

Li Yunchu: To build a world where robots can learn, the core is capturing the fundamental structure of problems. A key requirement for such worlds is consistency. This is where we strongly align with Marble—because we’re building a world consistent across space, time, viewpoints, and interaction modalities. Marble-generated space provides the indispensable foundational component. Imagine a robot pushing an object, only for it to vanish mysteriously—this plague many existing video prediction models. The robot receives no correct feedback signal. While the industry is deeply researching increasingly powerful video models, we believe our infrastructure can provide initial momentum, forming a data flywheel: transitioning from simulation-focused models to robot policy models, which execute tasks in the real world and collect new data, which then feeds back. This model needn’t be purely physics-driven or purely data-driven—it sits between both, capturing the essential problem structure while continuously expanding and evolving with accumulated data.

Host: I’ve collaborated closely with Li Feifei, who has a clear North Star guiding her. I wonder if you share a similar ultimate vision, or if you’re more pragmatic—focused solely on building systems that work and land?

Li Yunchu: My ultimate goal is to get robots working—truly performing tasks reliably in real environments. I’m very practical. During my postdoc with Li Feifei, we built a benchmark survey asking people what they’d most want robots to do. Among thousands of responses, one-third were related to cleaning. People simply dislike dirty, tedious work—exactly the scenarios we aim to solve with robotics.

Li Feifei: I deeply admire SceniX. While many robotics companies focus on models, Li Yunchu and his co-founders approach robotics with remarkable pragmatism. Though from academia, their instinct is to collaborate closely with real-world design partners and clients—in industrial labs, warehouses, or electronic component assembly lines. This approach is refreshingly novel and makes me incredibly excited about our collaboration.

Host: In my view, robotics must be precise—even if not perfect, it must be close enough. World Labs’ previous work in creative applications seems less demanding on precision; sometimes, slight inaccuracies or distortions are even considered artistic. From a technical standpoint, is combining these two approaches difficult? Or are they forever incompatible directions in design space?

Li Yunchu: In the long run, they will converge. Environment modeling doesn’t need to be perfect. The same applies to robotics models.

Host: Incidentally, is there a more formal term for “not perfect but close enough”?

Li Yunchu: Consider this example: throughout robotics application history, modeling has always been a cornerstone. Look at existing robots—aircraft, drones, vacuum cleaners, quadruped or biped robots—modeling has been the critical bridge enabling successful transfer from simulation to reality. For legged robots navigating snow or bushes, you don’t need a simulator that precisely replicates every bush or snowflake. You need a simulation system that captures the essential structure and enables systematic randomization testing in digital space—this is our goal. Thus, SceniX and World Labs are exploring the required fidelity level to model the vast physical world beyond robots, enabling seamless migration of digitally trained robot systems back to the real world.

5. The Unique Philosophy of Simulation: Counterfactual Reasoning, Reliability, and Speed Breakthroughs

Host: Some researchers (like Sergey Levine) argue simulations will inevitably diverge from reality, making real-world data collection indispensable. How viable is your approach of making simulation the foundation?

Li Yunchu: These aren’t mutually exclusive. Simulation is essentially predicting how the environment changes upon action—it is itself a world model. It doesn’t need to be purely physics-based; it can blend physical laws with data-driven learning. We continuously collect real-world data and use it, just with shifting emphasis at different stages of the data flywheel. Initially, we may emphasize physics to ensure correct consistency and structure, enabling world learning and robot policy training. As we accumulate more data through data collection and client partnerships, the model gradually shifts toward data-driven evolution. This transition, driven by data flow, perfectly combines the advantages of physics, geometry, and consistency with the power of data and compute.

Li Feifei: Let me add a philosophical perspective: the choice isn’t binary between “using simulation” or “not using simulation”—they work together to enable robot operation. Consider human intelligence: our brains constantly perform simulations. Why? Because simulation plays an irreplaceable role that real data cannot—counterfactual reasoning. You can mentally rehearse events that haven’t happened, couldn’t happen, or lack sufficient real-world data. Through rehearsal, you learn how to respond. Humans have always done this. Think of the World Cup—you’ve likely seen pre-game simulations in every match.

Whether digital or on a whiteboard, these simulations in preparation are fundamentally counterfactual reasoning. This is crucial for robotics because we simply don’t have enough real-world data. Take Waymo in autonomous driving—officially, they’ve used billions of hours of simulation data. In fact, Waymo relies more on simulation than real-world data. These are living examples—and remember, cars are among the simplest robots. Clearly, simulation plays a pivotal role in robot learning.

Li Yunchu: I’d like to add another point. Specifically, simulation delivers two core values: reliability and efficiency. On reliability: to ensure stable, robust operation in real environments, you need systematic coverage of all possible states and variables. Simulation allows systematic randomization of lighting, friction, geometry, object types, and physical parameters—ensuring comprehensive state space coverage, thus granting robots exceptional robustness.

The second aspect is efficiency. Many currently use teleoperation to collect data, but devices like exoskeletons often collect data slower than humans performing actions. Yet, for many of our clients, human speed is insufficient—they need to surpass it. Making robots faster isn’t merely about increasing motor voltage; gravity and other physical conditions remain unchanged. But in simulation, you can systematically accelerate training, allowing robots to consider all dynamic environmental changes. This is the efficiency gain simulation offers clients—whether in reliability or speed, simulation brings unique value.

Host: You’ve described your technical and platform capabilities. Could you detail current application scenarios?

Li Yunchu: There are two primary application areas: evaluation and training. First, evaluation—often overlooked in robotics. But if you’re tuning a robot model, you must clearly know its performance—this is the sole basis for iterative improvement.

Host: Yes. Incidentally, AI practitioners understand evaluation daily, but non-AI professionals often misunderstand it. Perhaps it’s worth clarifying what you mean by “evaluation.”

Li Yunchu: By evaluation, I mean accurately assessing a specific model checkpoint’s performance—e.g., success rate at 95% vs. 99.9%. The core industrial metric is time: how much actual work time does it take to distinguish between a 90% and a 92% success rate model? Doing this entirely in real environments would be excruciatingly slow. Current real-world robot evaluation iterates orders of magnitude slower than LLMs—not just due to task diversity.

Host: Exactly—because robots must take real physical actions, atoms must move in space.

Li Yunchu: Precisely—and must obey physical laws.

Li Feifei: Have you seen robot videos? Almost all are sped up 8x or 10x—because real-world execution is painfully slow.

Li Yunchu: Exactly. Real-world testing is not only lengthy but dangerous and expensive—iteration speeds are slowed by several orders of magnitude. Thus, many of our clients need a digital environment to evaluate their robot systems. Since our digital environment is validated to be highly consistent with the real world, events in simulation are highly likely to occur in reality. If a model checkpoint performs better in simulation, it will likely perform better in the real world—we’ve discussed this in our blog. This gives clients immense confidence to leverage digital signals for scalable, safe, and highly efficient evaluations.

This is the evaluation use case. Now, training. As I mentioned, control is key. You want precise control over states, parameters, lighting, friction, physical properties, object geometry, and types—all variations to cover complex scenarios and generate rich, informative data, enhancing robot robustness. Performing this exclusively in the real world is extremely difficult. Teleoperation data collection is slow, constrained by robot and device numbers, creating numerous operational challenges. But in simulation, everything is controllable and systematically configurable—you clearly know what data distributions you’ve covered, building confidence that robots will operate stably under those distributions. This confidence, efficiency, and scalability are precisely why clients choose our digital world for robot training.

Li Feifei: Ironically, even before deep discussions with SceniX, various Marble prospects had expressed demand for this—though we lacked capacity to serve them. We received calls from early-stage robotics companies, ranging from teams developing foundational models to those focused on downstream practical applications. We genuinely saw strong market demand.

6. Redefining Business Boundaries: Model-Agnostic, Embodiment-Agnostic Infrastructure

Host: When people hear you’re entering robotics, their first reaction might be that you’ll start printing 3D parts, coding robot brains, and assembling bodies—ultimately building a physical robot. But I believe that’s not your focus. Could you clarify which stage of the robotics R&D lifecycle your technology occupies? Where are your business boundaries, and where does the broader ecosystem pick up?

Li Yunchu: Think of what we’re building as infrastructure and accompanying software—people can use it to construct digital worlds where robots learn and evaluate. This infrastructure is inherently model-agnostic and embodiment-agnostic.

Host: I want to emphasize this point thoroughly. Although it seems obvious to you, it’s a subtle distinction for many: from your description, you’re not building robots—you’re creating an environment where other companies can place their robot brains to navigate and learn.

Li Yunchu: Exactly. Our clients use various robots—single-arm, dual-arm, fixed arms, mobile manipulators; end-effectors range from simple grippers to complex dexterous hands. Our platform is designed to naturally accommodate all forms, easily integrating different robot bodies into our generated digital world, empowering them with high reliability and efficiency to perform specific tasks in the real world.

Additionally, our platform is model-agnostic. We can directly use data generated in the digital world to train different models—either from scratch or for post-training on existing vision-language-action (VLA) or world-action models (WAM). For us, the specific model architecture matters little. Our core focus is providing a complete infrastructure and digital world, enabling robots to operate stably and reliably in real environments.

7. Practical Deployment Roadmap: From Semi-Structured to Ultimate Unstructured Environments

Host: You previously noted that market predictions for humanoid robots are overly optimistic—practical deployments are more likely in confined settings like warehouses in the near term. Could you elaborate on this, and how it impacts World Labs’ future business strategy?

Li Yunchu: Looking at the progression of robotics in real environments, it consistently follows a path from fully structured to semi-structured, finally approaching unstructured environments. Fully structured environments mean complete control and understanding of all configurations—like factories or car manufacturing lines, already highly automated decades ago.

Semi-structured environments involve partial control—like Amazon’s automated warehouses, or restaurants and hotels. Here, you intervene appropriately to reduce robot execution difficulty, but many objects remain uncontrollable—e.g., irregularly shaped clothing. Finally, unstructured environments—like homes. These represent the ultimate challenge.

Host: Especially mine—believe me, I have three dogs and a five-year-old child.

Li Feifei: Yes, and dogs.

Li Yunchu: Absolutely. Robustness stems from covering all possible scenarios robots might encounter. Thus, focusing on semi-structured environments first is far easier and more feasible than rushing toward unstructured ones. We will eventually reach unstructured environments, but we aim to do so sustainably and pragmatically.

Li Feifei: I believe your core point is that humanoids mimic the human body, and evolution optimized our form for unstructured environments. Our fingers and legs aren’t ideal tools for any single task. If our species’ sole survival goal were tree climbing, we wouldn’t have evolved our current anatomy—we’d likely have different fingers. But humanity evolved a remarkably general form, though not optimal for every task—precisely to survive in unstructured environments. From commercial and practical tech perspectives, however, unstructured environments and general bodies are among the hardest challenges—perhaps not even the right way to solve the problem. A better approach is specialization: using more customized bodies to solve narrower, specific problems. But SceniX’s challenge lies in ensuring its infrastructure is embodiment-agnostic, serving diverse robot bodies and semi-structured environments.

8. Commercial Outlook and Integration Strategy: Maintaining Just Right Rational Optimism

Host: Taking an economic perspective—comparing to LLMs—generative LLMs produce text or code ~10,000x faster than humans, at much lower cost. Economically, this makes perfect sense, as our brains aren’t efficient at these tasks. However, our brains and bodies are highly efficient at 3D navigation, moving in physical space, or picking up objects. So here’s the predictive question: Do you believe we could, in the foreseeable future, build robots matching human efficiency in daily, physically demanding labor (e.g., minimum wage jobs)? Is this 5 years away, or nearly impossible?

Li Yunchu: This will take a very long time. A robot operating in reality is a complex systems engineering endeavor—requiring meticulous coordination across hardware, software, algorithmic brain, down to details like finger friction coefficients. Realizing this requires considering countless factors and continuous iteration. But what excites me is being at the forefront of robot learning, constantly pushing technological progress. Today’s pace of advancement consistently exceeds my expectations. The projects I’m exploring now differ vastly from what I studied during my PhD. This shows the ecosystem is evolving rapidly—the various modules are progressively converging to build real robot systems. Yet we must maintain rational forecasting: while we’ll see great progress, reaching human-level efficiency and capability will still take considerable time.

Li Feifei: In today’s AI landscape, the hardest thing is maintaining just-right rational optimism. Even LLMs haven’t matched human brain efficiency—our brains operate on just 30 watts. We’re still far from that target.

Host: Still, in narrow, specific tasks like code writing or image generation, performance-to-power ratios may already be close. But in robotics, it’s vastly different. Has this changed your team’s strategic ambitions? Has it altered your development trajectory, or does it still align with your original entrepreneurial vision?

Li Yunchu: This has significantly reshaped our technical evolution path. Especially after collaborating with World Labs, we see enormous potential for higher efficiency and larger-scale expansion in the end-to-end environment modeling process. I’d also add that, despite LLMs’ impressive capabilities, you still wouldn’t blindly trust them to book flights or hotels—you’d want a human to review their output.

Moreover, robot models differ fundamentally from LLMs: once deployed, robot models must be ready-to-use and reliable in real environments. Yet we still lack sufficient data and infrastructure to enable robots to achieve such out-of-the-box reliability. Hence, building scalable digital worlds where robots can learn and evaluate unlocks massive potential—replacing expensive, unsafe real-world data collection with synthetic data generated in the digital realm, enabling large-scale learning and evaluation.

Host: I’ve seen many mergers like this—when philosophies align, operations usually run smoothly. But everyone wonders: Will you fully integrate the two systems immediately, or keep them relatively independent with a long-term plan over the next year? Li Feifei, how are you thinking about this—immediate deep integration, or treating it as a separate long-term project?

Li Feifei: A very insightful question. We’ve deeply discussed with team members like Changi, Sunonni, Justin, and Ben. Our current strategy is thoughtful and gradual—not rushing to merge codebases or teams. SceniX has a mature, relatively complete tech stack, customer base, and product form. We’re giving ample time for integration. Of course, we’ve already begun deep collaboration in simulation and action-conditioned foundational model development, and they were internal customers of Marble beforehand. Integration will happen, but we won’t rush to mix teams like tossing salad.

Host: How about location and office layout? Will the SceniX team relocate?

Li Feifei: Li Yunchu will move here. World Labs is now a cross-West Coast company, headquartered in San Francisco. I live in Palo Alto—feels like another state. Excitingly, we’ll open an office in New York, helping attract top East Coast talent. Additionally, we’ll equip offices on both coasts with robot hardware for testing and refining our engineering stack, enabling remote operation and debugging—essential capabilities for serving our clients.

Host: Let’s envision a specific scenario: What’s the ideal success state two years from now? What products will be launched? Who will be the customers? How will they use them?

Li Feifei: If we can validate landmark clients across several key vertical applications with the SceniX team, proving our system and infrastructure genuinely solve their automation needs, and these clients become beacon cases for scaling our business, we’ll be very satisfied.

Host: For robotics startups listening to this conversation, at what stage should they contact World Labs—early, or during mid-development?

Li Yunchu: For our current clients, since we deliver a Real-to-Sim-to-Real pipeline, simulation serves as the digital training and evaluation ground. Some clients need only the Real-to-Sim portion, wanting to digitize their tasks and evaluate robot systems within that space; others need the full pipeline to deploy trained policies onto hardware. Thus, our platform is highly flexible, customizable to client needs. Moreover, our partners are mostly nearing deployment, focused on very practical applications—where the robot solution creates immediate value, and these scenarios have dozens or even hundreds of replicable automation needs. We’ll work with World Labs to develop reliable solutions for these contexts. We’ve already showcased multiple real-world deployment scenarios in our blog posts, and we’ll further explore how to address key demands and constraints in actual deployment.

Host: Is it too early or too late for robotics companies to contact World Labs now?

Li Feifei: Not at all. We’re ready to collaborate anytime. Everyone is welcome to reach out—we’re eager to learn about your specific application scenarios.

Source: KeyPoints...