World Models Enter the Real Testbed: WorldArena 2.0 Challenge Launched

WorldArena 1.0 limitations

WorldArena 1.0 moved evaluation from visual appeal to practical usefulness by requiring physical consistency, controllability, 3D accuracy and embodied task functionality. The benchmark remained offline, focusing on video generation and simulation, and did not provide a systematic framework for real‑robot execution, online reinforcement‑learning (RL) loops, or multimodal perception under noise, latency and error accumulation.

WorldArena 2.0 upgrades

Version 2.0 expands three axes: modality (visual → visual‑touch), function (offline evaluation → online RL environment) and platform (simulator → real robots). The upgrade is organized as three tracks that progressively increase evaluation pressure.

Three tracks

Track 1 – Video Quality Evaluation on OOD Tasks

Track 1 retains the focus on video quality and physical consistency but adds longer prediction horizons, more complex tasks, and out‑of‑distribution (OOD) scenes and objects. The evaluation dimensions include visual fidelity, motion realism, content correctness, physical plausibility, controllability and 3D accuracy.

Track 2 – World Model as Online RL Environment

In Track 2 the world model becomes a closed‑loop RL environment: it receives policy actions, predicts the next observation, provides reward‑related feedback and supports repeated roll‑outs. A single prediction error propagates to subsequent actions, creating error accumulation that can destabilize learning. Existing experiments show modest policy gains over basic supervised fine‑tuning (SFT) but a clear gap to full simulators.

Track 3 – WAM on Real‑World Robotic Platforms

Track 3 moves evaluation to physical robots (e.g., AgileX, Franka) and tests task planning and action execution across multiple bodies. Two conditions are defined: vision‑only and tactile‑vision. The track stresses multimodal perception, long‑duration dynamics, robustness to sensor noise, latency, hardware variation and the ability to correct actions after deviations.

Results on the UniVTAC simulator show that tactile prediction achieved 21.26 PSNR and 0.746 SSIM, yielding 100 % success on the Insert HDMI task. The same models failed (0 % success) on the Lift Bottle task, which requires continuous force control; an ACT baseline reached 80 % success on Lift Bottle.

Multimodal alignment: visual, tactile, robot state and actions must be temporally and spatially consistent.

Sustained execution: the model must handle slip, friction and contact forces over long horizons, not merely detect instantaneous contact.

Closed‑loop correction: after a prediction or execution error, the policy must use new visual/tactile feedback to re‑align the trajectory.

Real‑world robustness: the system must tolerate sensor noise, latency, device differences and material variations while maintaining stable operation.

Key dates: challenge opens 10 July 2026, final submission deadline 30 August, results announced 15 September, awards 27 September.

Resources: http://iros2026challenge.world-arena.ai/, https://physical-world-models.github.io/IROS2026, http://v2.world-arena.ai, https://arxiv.org/abs/2605.17912, https://github.com/WorldArena2/WorldArena-2.0, https://huggingface.co/spaces/WorldArena/WorldArena2.0