Fei-Fei Li's World Labs launches R2S2R engine to train real robots in simulation

World Labs, the spatial intelligence company founded by Fei-Fei Li, has released a new robot policy training and evaluation engine called R2S2R (Real-to-Sim-to-Real), built with the help of newly acquired robotics startup SceniX. The engine extends World Labs' spatial intelligence ambitions from generating interactive virtual worlds to training, evaluating and deploying real robots — closing the loop from simulation training to real-world operation for the first time.

R2S2R has two halves. Real-to-Sim moves real robots, sensors, objects and interactions into simulation, building virtual environments aligned with real tasks; Sim-to-Real trains and evaluates policies in those virtual environments, then deploys the resulting strategies back into the physical world.

Li said on X that the results are striking: Policies trained with this method can run autonomously for a full hour without human intervention, even though they have never used real-world data. The work advances what Li has called the most critical piece of her world-model taxonomy: the simulator, as opposed to the renderer that generates observations and the planner that outputs actions.

World Labs explains the motivation in a blog post: the main bottleneck for robotics today is no longer just model architecture, but the lack of experience and evaluation that can be acquired at scale. Unlike internet data for large models, every piece of robot experience must actually happen — each experiment consumes hardware, manual resets and failure recovery, making it hard to systematically cover diverse environments, object properties and failure conditions.

In the Real-to-Sim phase, the team captures robots, sensors, environments, objects and task demonstrations, then rebuilds them into an interactive virtual world with a generative world-modeling system — reproducing not just appearance but physical and dynamic properties as closely as possible. The team then runs identical action sequences in both simulation and reality, comparing observations, object reactions and outcomes to verify alignment.

In the Sim-to-Real phase, the aligned simulation becomes a scalable training and evaluation engine: robots learn new policies in simulation, the evaluation system actively hunts for failure states, generates new interaction experience around those weaknesses, and only the most promising policies are deployed to real hardware.

According to the report, policies trained purely in simulation transferred directly to robot platforms including ALOHA, RB-Y1, YAM, Flexiv and xArm, completing tasks such as bin packing, wire winding, test-tube transfer and single-object grasping in cluttered scenes. Several tasks ran autonomously for one hour with no failures and no human takeover. In an ALOHA dual-arm block handover task, the team tested GR00T N1.6 and π₀.₅ policy architectures with 2,000 simulation trials and 100 real-robot trials per checkpoint, finding that checkpoints that performed better in simulation generally performed better on real hardware.

A week before the R2S2R launch, World Labs announced it had acquired SceniX, a startup focused on robot simulation, training and evaluation. SceniX co-founder Li Yunzhu is an assistant professor at Columbia University who did postdoctoral research with Fei-Fei Li at Stanford. In an a16z interview, Li said the acquisition does not mean World Labs will build robots itself — the goal is infrastructure decoupled from robot form factors and model architectures.

The two teams plan to integrate simulation, foundation models and action-conditioned models over time, starting with real-world deployment validation with customers in semi-structured settings such as warehouses, labs and electronics assembly lines.