Fei-Fei Li on Robotics' Hardest Problem: Teaching Machines the Physical World
Key points
Key takeaways from the a16z interview with World Labs co-founder and CEO Fei-Fei Li, hosted by Andreessen Horowitz's Martin Casado (August 2026):
The hardest problem in robotics is the physical world. Li frames spatial intelligence — giving machines a true understanding of, and ability to interact with, the physical world — as one of AI's biggest unsolved problems, and the throughline connecting her work from ImageNet to World Labs.
World Labs has acquired SceniX. The interview marks the company's acquisition of SceniX, uniting two teams working on machine understanding of the physical world; SceniX co-founder Yunzhu Li, an assistant professor at Columbia University, joins the conversation.
Data is the bottleneck in robotics. Unlike language models trained on abundant internet text, robotics lacks the vast, diverse interaction data needed to learn, making the data bottleneck the central constraint on building capable robots.
Simulation could unlock the next generation of robots. Li argues that real-to-sim-to-real pipelines — capturing the real world, generating varied scenarios in simulation, then transferring back — are essential to generating the scale and diversity of experience robots need to learn.
Training robots differs fundamentally from training language models. The conversation contrasts video models with real-to-sim-to-real approaches and argues that a robotics foundation model may require a fundamentally different paradigm than the text-pretraining recipe behind large language models.
Evaluation and real customer use cases matter as much as training. Beyond model building, the discussion covers how to evaluate robot performance, generate synthetic data, and ground the work in real deployments rather than benchmarks alone.
Humanoids and semi-structured environments are the grand challenge. Li and Casado discuss integration plans following the SceniX deal and what success looks like in two years, with humanoids operating in messy, semi-structured human environments framed as the field's grand challenge.
Transcript unavailable — summary based on title and description only.
Read more: a16z · World Labs
