CVPR 2026 WorldArena Challenge Launches with Amap’s Open‑Source High‑Performance World Model Baseline

WorldArena Challenge: The "Big Test" for World Models

WorldArena Challenge, hosted by the CVPR 2026 Video World Model Workshop, is an international competition jointly organized by Amap CV Lab, Manifold.ai, Tsinghua University, Princeton University, NUS, HKU and other leading academic institutions.

Unlike previous world‑model evaluations, the challenge focuses on "real‑world functionality". It assesses not only visual fidelity but also whether generated videos obey physical laws and support robotic manipulation. The evaluation uses the WorldArena Benchmark, developed by eight top universities, covering 16 core metrics and three real‑world application tasks.

Track 1 measures comprehensive perception quality across six dimensions—visual quality, motion quality, content consistency, physical‑law compliance, controllability, and 3D accuracy—through 16 quantitative indicators, aggregated into an EWMScore for ranking.

Track 2 evaluates embodied‑task functionality, examining three uses of world models: data synthesis for policy improvement, policy evaluation as a physics‑simulator substitute, and direct action planning. Top solutions may receive additional weighting for planning capabilities.

Getting Started: High‑Performance Open‑Source World Model

Amap has fully open‑sourced its world model ABot‑PhysWorld , which currently leads the WorldArena leaderboard. ABot‑PhysWorld is the first sub‑project of the upcoming ABot‑World series and targets physical‑consistent video generation for embodied scenarios.

Key technical breakthroughs include:

Four‑dimensional Generalization Data : Cleaned from 3 M raw samples to 300 k high‑quality SFT data, covering robot morphology, 50+ task types, 10+ scenes, and 1 000+ object categories.

DPO Preference Alignment : Constructed 10 k preference pairs with a VLM‑as‑Judge and applied Direct Preference Optimization to correctly choose between physically correct and incorrect outcomes, markedly reducing clipping and deformation.

Dense Action Map Fine‑Control : 110 k action‑control samples encode robot motions as dense spatial signals; Context Blocks fuse these with video latents for precise motion injection.

On the independent PAI‑Bench, ABot‑PhysWorld achieves a 0.8491 overall score and a 0.9306 domain score, surpassing models such as GigaWorld, Wanx‑2.5, Veo 3.1, and Sora 2, and breaking the long‑standing trade‑off between visual quality and physical compliance.

ABot‑PhysWorld also ranks at the top of the WorldArena leaderboard, demonstrating Amap’s technical strength in embodied world modeling. To ensure fairness, the model is excluded from final prize evaluation, and all weights, training code, and data pipelines are publicly released for fine‑tuning and innovation.

Competition Schedule and Participation Guide

The prize pool exceeds $14 000, with first, second, and third places in each track. Winners receive a CVPR workshop presentation slot, and top teams may earn cross‑track awards. Submission channels are open, supporting real‑time leaderboard updates. The final submission deadline is May 25 2026, results are announced on June 1, and the award ceremony occurs on June 4 during CVPR.

Submission workflow (approximately half a day):

Prepare data: download val_dataset / test_dataset from Hugging Face.

Generate video: output a ≥640×480, 121‑frame @24 fps video (initial frame + text/action) with your model.

Package and submit: zip the video folder with model_README.md and upload via the official website or email.

For more details, resources, and registration, visit the official website and GitHub repositories linked in the announcement.