Robots trained in simulation alone
On 28 July 2026 World Labs showed first results of its R2S2R engine: a real task is rebuilt as a simulation, policies trained in it with no real data run on several robots, and the simulation predicts which policy is better on hardware.
Why it matters
The bottleneck of robotics here is not the policy architecture but the amount and variety of experience; the company proposes to get it from a simulation aligned with reality. An editorial assessment: everything is the company’s own result, without a paper or code, so “for the first time” and “a new bar” remain its words.
By the post, the SceniX engine captures a real task (robot, sensors, environment, objects, demonstrations), rebuilds it as an interactive world and varies appearance, placement, clutter, physics and camera. Validation: the same action sequence run in simulation and reality, with observations and outcomes compared. Claimed: a policy for ALOHA (two-handed box packing) trained only in simulation transferred to the robot; five tasks on four further platforms (RB-Y1, YAM, Flexiv, xArm) ran autonomously for an hour each without intervention. Evaluation: on an ALOHA cube-handover task each checkpoint was tested on 2,000 simulated trials (1,000 in-distribution and 1,000 outside it) and 100 real ones (50 and 50); the company says policy rankings in simulation hold on hardware and that a simulation need not match real success rates. What the record does not state. The text of the post has no success rates and no rank-correlation figure, since the result is shown as a chart; it does not name the architectures compared (The Decoder of 15 August writes of GR00T N1.6 and pi-0.5, which the post’s text does not contain). There is no paper, code or dataset. “Zero real data” and “for the first time” are the company’s words; transfer from simulation to a robot has long been known, and the record does not say it is new.