Fei-Fei Li on Spatial Intelligence and Robotics
World Labs, founded by Fei-Fei Li, announced the acquisition of Cynics, a robotics simulation company co-founded by Yun-Zhu Li. The two organizations are combining spatial intelligence and world models with practical robotics expertise to develop a real-to-sim-to-real pipeline that enables robots to train and evaluate in digital environments before deployment in physical spaces.
Summary
Fei-Fei Li and Yun-Zhu Li discuss World Labs' acquisition of Cynics and their shared vision for advancing robotics through spatial intelligence and world modeling. World Labs has been building large generative world models, with their primary model Marble capable of generating geometrically consistent 3D worlds from text, image, or multi-image prompts. Cynics specializes in creating dense reconstructions of real environments and simulating their dynamics, enabling a real-to-sim-to-real pipeline for robotics training and evaluation.
The core problem Cynics addresses is the scarcity and inefficiency of real-world robotic data collection. Unlike language models trained on abundant internet data, robotics companies struggle with limited data, slow iteration cycles, and the high cost and danger of physical testing. By digitizing real environments with sufficient fidelity, Cynics enables scalable training and rapid evaluation of robotic systems in simulation before real-world deployment.
Fei-Fei emphasizes that simulation's critical role is enabling counterfactual reasoning—allowing systems to learn from scenarios that haven't occurred or cannot be easily replicated in the real world. She cites examples like Waymo's use of billions of hours of simulation, demonstrating that even for relatively simple robots like self-driving cars, simulation is essential. Yun-Zhu highlights two key benefits simulation provides: reliability through systematic coverage of state spaces and variations, and efficiency by enabling faster-than-human speed training through controlled parameter manipulation.
Regarding the fidelity required, they argue that perfect physics simulation is unnecessary. Instead, models must capture the essential structure of problems while allowing for domain randomization. The approach combines physics-based and learning-based modeling, starting with stronger physics emphasis for consistency, then transitioning toward more learning-based methods as real-world data accumulates.
The acquisition brings complementary technical strengths: World Labs contributes generative modeling and 3D reconstruction capabilities, while Cynics brings full-stack robotics expertise, robotics-specific research leadership, and simulation technology. The team includes Changxi Zheng (Columbia professor and former VFX engineer), Sunny Hu (engineering leader with Amazon background), and other talent pools previously absent from World Labs.
On application strategy, they position their work as embodiment-agnostic and model-agnostic infrastructure. Their customers deploy various robots (single arms, bipedal systems, mobile manipulators, different grippers), and the platform integrates different embodiments into digitalized worlds. Rather than building specific robots, they're building environments where any robot can learn and evaluate.
Regarding timeline and scope, Yun-Zhu advocates a pragmatic approach, focusing initially on semi-structured environments (warehouses, restaurants, controlled settings) before progressing to unstructured home environments. He references survey data showing one-third of thousand public requests for robotic tasks involve cleaning—"dull and dirty" work people want automated. They're working with customers close to deployment stage on practical, valuable tasks.
On power efficiency and long-term feasibility, Yun-Zhu expresses measured optimism. While confident in rapid progress, he cautions that achieving human-level efficiency in robotics will take considerable time, unlike narrow AI tasks where performance-to-power ratios may approach human capability. He emphasizes that successful robots require thoughtful system integration across hardware, software, and numerous physical parameters.
The companies plan a measured integration approach: maintaining Cynics' fairly contained tech stack and customer relationships while collaborating on simulation and foundation models. Cynics is already using Marble as an internal customer. Yun-Zhu is relocating from New York to San Francisco, with World Labs establishing a bi-coastal presence to attract East Coast talent and test robotics infrastructure remotely.
About this episode
Last week, World Labs announced its acquisition of SceniX, bringing together two teams working on one of AI's biggest unsolved problems: how to give machines a true understanding of the physical world. Martin Casado sits down with Fei-Fei Li, co-founder and CEO of World Labs, creator of ImageNet, and pioneer of spatial intelligence, alongside Yunzhu Li, co-founder of SceniX and assistant professor at Columbia University. They discuss why World Labs acquired SceniX, how simulation can unlock the next generation of robotics, and why training robots may require a fundamentally different approach than training language models. The conversation explores real-to-sim-to-real pipelines, world models, robotics foundation models, evaluation, synthetic data, and why the future of AI depends not just on understanding language—but on understanding and interacting with the physical world.
Key Insights
- Fei-Fei Li argues that simulation's critical advantage over real-world data is enabling counterfactual reasoning—allowing AI systems to learn from scenarios that haven't occurred, cannot be easily replicated, or lack sufficient real-world examples.
- Yun-Zhu Li contends that perfect physics simulation is unnecessary for robotics; instead, models must capture the essential structure of problems while using domain randomization to ensure robots generalize across variations.
- The speakers argue that robotics iteration speeds are multiple orders of magnitude slower than language model iteration because physical robots must obey laws of physics and move through actual space, making evaluation in real environments prohibitively costly and time-consuming.
- Cynics and World Labs position their platform as embodiment-agnostic and model-agnostic infrastructure, meaning it can integrate different robot bodies and train different types of models, rather than building specific robotic systems.
- Fei-Fei Li claims that simulation provides two distinct benefits: reliability through systematic coverage of state spaces and variations, and efficiency by enabling faster-than-human-speed training through controlled parameter manipulation in the digital environment.
- Yun-Zhu Li asserts that robotics applications historically progress from fully structured (factories) to semi-structured (warehouses) to unstructured environments (homes), and their pragmatic approach focuses on semi-structured environments as the current frontier rather than pursuing humanoid robots for unstructured home automation.
- The speakers argue that robots functioning reliably in real environments require out-of-the-box reliability unlike language models, which typically need human oversight; this fundamental difference makes scalable training data generation through simulation particularly valuable for robotics.
- Yun-Zhu Li expresses measured optimism that while robotics will see rapid progress, achieving human-level efficiency in 3D navigation and manipulation tasks will require considerable time, noting that even human brains (30 watts) vastly outperform current AI in power efficiency for these tasks.
Topics
Transcript
We are building the next frontier of AI, which is what we call spatial intelligence. At Cynics, we are developing what we call a real-to-sim-to-real pipeline. We can replace all the data, all the evaluation we need in the real environment by using the data that can generate at a scalable way in our digital world. Think about human intelligence. We do a lot of simulations in our head. You know why? There's a very important role simulation plays that real world data doesn't play, which is counterfactual reasoning. What we are building is a consistent world. Consistence both over space, over time, over different viewpoints, and over different type of interactions. My North Star is I want the robot…
Full transcript available for MurmurCast members
Sign Up to AccessMore from The a16z Show
Steven Sinofsky: AI Doesn't Need New Rules Yet
Steven Sinofsky argues that AI regulation is moving too fast and without sufficient understanding of the technology, drawing parallels to how previous innovations like cars and the internet evolved before being regulated. He advocates for open source AI development and warns against regulatory capture, where companies lobby for rules that eliminate competition rather than protect the public.
Ben Horowitz: The Fight Over Open Source AI
Ben Horowitz argues that open source AI is critical for safety, competition, and innovation, comparing it to successful open source technologies like Linux and the internet. He contends that banning open source would harm academia, stifle competition, and create dangerous monopolies, while open models allow distributed security improvements and prevent single-company control of AI.
Sriram Krishnan on Open Source AI's Biggest Week Yet
Sriram Krishnan, former White House AI policy advisor, discusses how recent open-source AI models like Kimi K3 are reshaping the industry by increasing competition, creating pricing pressure on frontier labs, and raising important questions about security, distillation, and America's competitive position in the global AI race.
Building the Physical AI Stack | Travis Kalanick on TBPN
Travis Kalanick discusses his new company Atoms, which is building industrial AI solutions to automate mining, food production, and transportation. He raised $1.7 billion and explains how autonomous systems are increasing productivity and safety while reducing operational costs across these industries.
Travis Kalanick Is Back | Building the Future of Industrial AI
Travis Kalanick discusses his return to entrepreneurship with Atoms, a conglomerate applying industrial AI to transform manufacturing, logistics, and autonomous systems across food, mining, and transport industries. He reflects on lessons from Uber's 2017 crisis, his stealth-mode building strategy, and the framework of 'atoms-based computation' as the next frontier beyond software.