WorldArena, the benchmark that has become a reference point for world models, opened its 2.0 challenge at IROS 2026 with two new tracks built to test one thing the field dodges: can the model actually drive a robot, or only generate a pretty video?
Track 1, WorldIncept, sets 70 formal tasks, 50 clean in-distribution scenes on a white table and 20 out-of-distribution scenes on coloured or textured tables. Every model is tested on the same dataset, and a single prediction error compounds across rounds of interaction, so the model must stay stable and controllable and reflect state changes truthfully enough to reward the policy.
Track 2, MW2, runs on the RoboTwin environment with the “Adjust Bottle” task, pushing fine pose control and grasp stability. Track 3, ViTacX, scores on real-robot execution success within a step budget and without triggering safety intervention, weighting vision and visuo-tactile directions.
The organising team, led by Tsinghua with Peking, CMU, Stanford, Princeton, the University of Hong Kong, NUS and the Chinese Academy of Sciences’ automation institute, framed the additions as a direct reply to the “sim-to-real gap” critics raised. Track 3 is the honest version: a model that looks good in simulation must now prove it on metal.
The result worth noting is which route won. The field now spans four technical paths, JEPA, spatial or structured 4D, and pixel-generation among them, and the JEPA-route teams took the stronger positions, Shiqi Weilai first in Track 1 and Chenhunxian second in Track 2. The hosts read that as a signal, not a verdict, that prediction stability may matter more than visual fidelity when the model has to act.
The team also guards against a new failure: WorldArena becoming a target the field optimises against, trading real progress for benchmark scores. Daily and challenge leaderboards stay separate, and updates run on a fixed cadence so the metric cannot quietly become the product.
Editor’s note: This is an adapted translation of the original LeiPhone report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.leiphone.com/category/private/A0uKolgD2esWoUfT.html.