Fei-Fei Li on Spatial Intelligence and Robotics

What It Covers

Following a16z's Martin Casado, Fei-Fei Li of World Labs, and Yunzhu Li of Scenix discuss the acquisition of Scenix and the case for spatial intelligence as the next AI frontier — using simulation, 3D world models, and real-to-sim-to-real pipelines to solve robotics' core data and evaluation bottlenecks.

Key Questions Answered

• •Real-to-Sim-to-Real Pipeline: Robotics development is bottlenecked by slow, costly real-world data collection and evaluation. Scenix's approach maps physical environments into geometrically aligned digital worlds, enabling training and evaluation entirely in simulation. If a robot checkpoint performs better in the digital environment, it reliably performs better in the real world — dramatically accelerating iteration cycles.
• •Simulation's Unique Role — Counterfactual Reasoning: Simulation does not simply replace real-world data; it enables scenarios that cannot be collected physically. Like athletes reviewing game plans on whiteboards, robots need exposure to edge cases and rare states. Waymo publicly reports using billions of simulation hours, and cars represent the simplest two-dimensional robot — underscoring simulation's necessity for more complex tasks.
• •Evaluation Speed as the Hidden Bottleneck: Distinguishing between a robot policy checkpoint performing at 90% versus 92% success rate requires extensive real-world trials, taking orders of magnitude longer than evaluating language models. Simulation with proven real-world alignment compresses this process, giving robotics teams the fast feedback loops that software AI teams take for granted during model development.
• •Embodiment-Agnostic Infrastructure Over Hardware: World Labs and Scenix are building model-agnostic and embodiment-agnostic infrastructure — not robots. Their platform accepts single arms, bimanual systems, mobile manipulators, and varied end effectors. Robotics companies at near-deployment stages can plug their hardware and existing foundation models into the simulation stack for training data generation and policy evaluation without switching embodiments.
• •Semi-Structured Environments Before Humanoids: Robotics deployment follows a progression from fully structured environments like car factories, to semi-structured environments like Amazon warehouses, to unstructured home environments. Humanoid bodies are evolutionarily optimized for generality, not efficiency at specific tasks. Near-term commercial robotics value lies in semi-structured industrial and logistics settings, where constrained variability makes simulation coverage tractable and reliability achievable sooner.