The final standings of the WorldArena 2.0 Challenge, held with IROS 2026, show different world-model approaches winning different kinds of tests. WorldArena is a benchmark built by a Tsinghua University team with universities in China and abroad, and the 2.0 edition extends it from vision to visuo-tactile input, from offline tasks to online reinforcement learning, and from simulation to real robots, Leiphone reported.
The challenge ran three independently ranked tracks with a combined prize pool of $7,000, and final results were published on September 16.
Track 1 scores video prediction with 15 metrics across 70 tasks, 50 in-distribution scenes on white tabletops and 20 out-of-distribution scenes with colored or textured tables. Of 81 ranked models, Visincept's WorldIncept placed first with 72.33, ahead of BWM-Turbo from Tongji University's spatial intelligence team at 71.49. The margin was just 0.84 points. WorldIncept led on physics adherence, 3D accuracy and semantic alignment, while BWM-Turbo ranked first on background consistency and JEPA similarity. Visincept explicitly follows the JEPA approach.
Track 2, new this year, uses a world model as the environment for online reinforcement learning on a RoboTwin 2.0 bottle-adjusting task, then tests the trained policy in simulation. Zhongguancun Academy's MW2 won with 72.93 among 47 models, 0.53 points ahead of TTWM, a competition version of a goal-conditioned causal world model from another JEPA-focused startup. The runner-up's team said training in its model raised a pi0.5 policy's success rate from 55.46% to 72.40%.
Track 3, also new, runs on real AgileX dual-arm and Franka Panda robots. Its visuo-tactile tasks include picking up chips, peeling a cucumber and inserting a plug; its vision-only tasks include wiping tables, pouring water and folding clothes. ViTacX from Xiaomi's MiRobot team won overall with 76.64. Its lead came from the vision-only sub-ranking, which it topped with 85.00, including perfect scores on three tasks. On the visuo-tactile sub-ranking it placed third with 66.67, while ShanghaiTech University's OmniFlow ranked second on both.
Organizers noted that participation was optional for each track, so no single model has yet been shown to lead on video generation, policy training and real-robot manipulation together.
SEE ALSO: Xiaomi Open-Sources Robotics-U0 Embodied World Model and Training Stack · Muka Robotics Charges WorldArena With Its LJM World Model