A research collective including Yilun Du, Lvmin Zhang (creator of ControlNet), and Haotian Zhang has released 'Training Object Permanence in World Models' (arXiv: 2609.28654, Project: object-permanence.world). While video generative models produce photorealistic scenes, they routinely violate physical object permanence—causing occluded objects to magically vanish, morph, or teleport. The researchers introduce WROP, a cognitive benchmark spanning 150 tasks and 1.5 million training trajectories, alongside PWM-WROP, an open-source 16B world model trained on AWS Trainium2 that secured top rank among video continuation models in blind Elo evaluations.

Key Takeaways

  • ✓Addresses foundational physical hallucinations in video generators where occluded objects disappear or deform.
  • ✓Releases WROP, a data infrastructure spanning 150 programmatic Blender cognitive tasks and 1.5M synthetic samples.
  • ✓Open-sources the 16B-parameter physical world model PWM-WROP alongside native PyTorch training recipes for AWS Trainium2.
  • ✓Outperforms 14 competitive video generation models, ranking first among video continuation architectures in blind Elo trials.
  • ✓Interactive project portal, evaluation exams, model checkpoints, and datasets released openly at object-permanence.world.
🔗

Project Links & Resources

Direct Access
Direct access to official project resources and documentation
📝
WROP 官方项目主页与交互式演示
www.object-permanence.world/
↗
🐙
Object-Permanence GitHub 官方开源仓库
github.com/hokindeng/object-permanence
↗
🌐
ArXiv 论文预印本 (2609.28654)
arxiv.org/abs/2609.28654
↗
📝
Hugging Face Daily Papers 讨论主页
huggingface.co/papers/2609.28654
↗
🐙
GitHub: hokindeng/object-permanence%60
github.com/hokindeng/object-permanence`
↗
🔬

In-Depth Technical Analysis

Core Background & Industry Pain Points While video generative architectures achieve remarkable photorealism, they fundamentally struggle with physical common sense. Cognitive hallmarks like object permanence and solidity—intuitive even to human infants—are frequently violated: objects occluded behind barriers morph, vanish, or teleport, crippling the reliability of video models as world simulators for embodied robotics and autonomous driving. ### Architecture Highlights & Internals The research group introduces WROP (World Reasoning with Object Permanence), featuring 150 procedurally generated Blender tasks spanning six core cognitive domains. Generating over 1.5 million synthetic trajectories with randomized camera angles and lighting, the framework isolates physical causality from visual noise. Crucially, the team pre-trained PWM-WROP, a 16B-parameter physical world model implemented in native PyTorch on AWS Trainium2 clusters, designed to preserve long-horizon geometric permanence through occlusion boundaries. ### Authoritative Benchmarks & Measured Scores Tested on a standardized 300-question cognitive examination across 14 frontier video architectures, PWM-WROP secured first place among video continuation models in blind pairwise Elo evaluations. On extended multi-second occlusion sequences, trajectory coherence reached 91.8%, slashing spatial deformation errors from 64.3% down to 12.4%. ### Developer Hands-on Guide Researchers can explore task visualizations at https://www.object-permanence.world/, clone the open-source PyTorch codebase from GitHub (hokindeng/object-permanence), and download model weights from Hugging Face.