AnswersPublished16 sources
WorldArena 2.0’s results show three different tests, and three winners
WorldArena 2.0 had three separate track winners: Visincept’s WorldIncept, Zhongguancun Academy’s MW2, and Xiaomi Robotics’ ViTacX. The expanded benchmark covered video prediction, world models used for online reinforcement learning, and real robot manipulation—so each track measured a different capability.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- WorldArena 2.0 had three separate track winners: Visincept’s WorldIncept, Zhongguancun Academy’s MW2, and Xiaomi Robotics’ ViTacX.
- The expanded benchmark covered video prediction, world models used for online reinforcement learning, and real robot manipulation—so each track measured a different capability.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What were the final results of the WorldArena 2.0 Challenge held with IROS 2026, including its Tsinghua-led benchmark origins and expansion. Article summary: The reported winners of the WorldArena 2.0 Challenge at IROS 2026 were Visincept’s WorldIncept in Track 1, Zhongguancun Academy’s MW2 in Track 2, and Xiaomi MiRobot’s ViTacX in Track 3; these were separate track titles, . Topic tags: general, general web, user generated, academic. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, chart
WorldArena 2.0 put embodied world models through three different tests: predicting video, supporting online reinforcement learning, and manipulating objects with real robots. Each track had its own winner, so the results are best read as three separate titles—not as a single overall championship. 2
19
The final leaderboard was announced on September 16, 2026. The challenge had three independently ranked tracks and a combined prize pool of $7,000. 18
19
| Track | What it tested | Winner |
|---|---|---|
| Track 1 | Video prediction | WorldIncept, from Visincept |
| Track 2 | World models as environments for online reinforcement learning | MW2, from Zhongguancun Academy |
| Track 3 | Real-robot manipulation | ViTacX, from Xiaomi Robotics |
A broader test for embodied world models
WorldArena began as a benchmark developed by a Tsinghua-led team with other universities and research institutions. It was designed to assess both what embodied world models can perceive and how useful they are for tasks. WorldArena 2.0 widened that scope in three ways: from vision to vision-and-touch inputs, from offline evaluation to online reinforcement learning, and from simulation to real-robot tasks. 19
27
Those expansions are reflected in the challenge’s three tracks: video prediction, the use of a world model in online reinforcement learning, and real-robot manipulation. 2
19
Track 1: WorldIncept edges out BWM-Turbo
Track 1 evaluated video prediction using 15 metrics across 70 tasks: 50 scenes with white tabletops and 20 with colored or textured tables. Reports list 81 models in the rankings. 16
19
WorldIncept scored 72.33, while Tongji University’s BWM-Turbo came second with 71.49—a gap of 0.84 points. 16
19 WorldIncept led in physics adherence, 3D accuracy, and semantic alignment. BWM-Turbo performed particularly well on background consistency and representation similarity.
17
The two models also took different technical approaches. Reports describe WorldIncept as following the JEPA approach, which emphasizes learning representations in latent space. BWM-Turbo uses a diffusion-transformer architecture with dynamic memory. 2
17
Track 2: MW2 wins the online-learning test
Track 2 asked whether a world model could serve as an environment for online reinforcement learning, an evaluation area added in WorldArena 2.0. 2
19 Zhongguancun Academy’s MW2 won with 72.93 points. TTWM placed second with 72.40, a difference of 0.53 points.
17
Coverage describes TTWM as a goal-conditioned causal world model from Chenhunxian Technology and associates the team with the JEPA approach. 2
17 The result reflects performance in this specific track; its score alone does not show that a model is better across every robot-learning task.
Track 3: ViTacX takes the real-robot track
Xiaomi Robotics’ ViTacX won Track 3, which evaluated real-robot manipulation. The broader WorldArena 2.0 benchmark also expanded to include vision-and-touch capabilities. 18
19
27
That victory answers a different question from the video-prediction and reinforcement-learning results: ViTacX topped the real-robot track, but did not thereby win the other two. 2
19
Three wins, not one overall ranking
The tracks measure distinct capabilities: predicting video, using a world model in online reinforcement learning, and carrying out tasks with real robots. Different winners are therefore not contradictory. The results show different models leading in different evaluations; they do not identify one model as champion across all three. 2
19
When comparing the leaderboard, the task matters as much as the score. A win in one track is not, by itself, evidence that the same model would lead in another track or in tasks that were not evaluated.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "WorldArena 2.0’s results show three different tests, and three winners"?
WorldArena 2.0 had three separate track winners: Visincept’s WorldIncept, Zhongguancun Academy’s MW2, and Xiaomi Robotics’ ViTacX.
What are the key points to validate first?
WorldArena 2.0 had three separate track winners: Visincept’s WorldIncept, Zhongguancun Academy’s MW2, and Xiaomi Robotics’ ViTacX. The expanded benchmark covered video prediction, world models used for online reinforcement learning, and real robot manipulation—so each track measured a different capability.
Sources
163.comJEPA、空间智能路线正面交锋：WorldArena 2.0 挑战赛开考「能不能被机器人真正用起来」
qifukexue.comWorldArena 2.0挑战赛结果揭晓：JEPA路线视启未来与晨昏线分获两赛道第一，真机闭环成世界模型新考
e-acc.ai全部 AI 动态 · 第 2 页
ar5iv.labs.arxiv.orgWorldArena 2.0: Extending Embodied World Model Benchmarking on Modality, Functionality and Platform
x.comWorldArena 2.0 at IROS 2026: Visincept, Zhongguancun Academy and Xiaomi Win Its Three Tracks