Resources for world modeling.

Provides remote, validated physical experiments and recorded action-state feedback for wireless-network world-model development; an experimentation platform rather than a completed autonomous world model.

arXiv Source ↗

VidaForge

Inspectable video-data recipes with provenance and controlled pretraining studies on Wan 2.1 and V-JEPA 2.1; adjacent infrastructure for predictive pretraining.

arXiv Source ↗

UnifoLM-WMA-0

Unitree framework that predicts future robot interactions for visual simulation and policy enhancement across embodiments.

Project GitHub Source ↗

NVIDIA Cosmos

World foundation model platform for Physical AI (robots + AD); open-weight under permissive license

GitHub arXiv Project Source ↗

NVIDIA Cosmos 3

Open omnimodal WFM unifying reasoning, world generation, simulation, and action modeling

arXiv Report GitHub Project Source ↗

NVIDIA FlashDreams

High-performance inference and serving library for interactive autoregressive video and world models

Project GitHub Source ↗

NVIDIA Cosmos-Predict2.5

Next-gen Cosmos WFM: flow-based, unifies Text/Image/Video2World; open checkpoints

arXiv GitHub Source ↗

stable-worldmodel

Reproducible world-model research platform with data layer, baselines, planners, and OOD tasks

arXiv GitHub Source ↗

minWM

Full-stack framework for building real-time interactive video world models from open video backbones

arXiv GitHub Source ↗

Nano World Models

Minimalist future-video-prediction codebase with configs, eval scripts, and checkpoints

arXiv GitHub Source ↗

SANA-WM

Open minute-scale, 720p video world model with 6-DoF camera control

arXiv Project GitHub Source ↗

Causal Forcing

Open recipe for real-time autoregressive video diffusion world rollouts

arXiv Project GitHub Source ↗

Helios

Open real-time long-video generation stack relevant to live world modeling

arXiv Project GitHub Source ↗

OpenDWM

Open-source toolkit for driving world models (SenseTime)

GitHub Source ↗

Matrix-Game

Open interactive game world model stack (SkyworkAI)

GitHub Source ↗

Genie Envisioner

Open robotic manipulation world foundation platform

arXiv Project GitHub Source ↗

AgiBot World

Large-scale manipulation platform and dataset for embodied world models

arXiv Project GitHub Source ↗

Gaussian World Model

Gaussian world model codebase for robotic manipulation

arXiv Project GitHub Source ↗

HY-World 2.0

Open 3D world generation / simulation stack (Tencent)

GitHub Source ↗

HunyuanWorld 1.0

Text/image-to-3D explorable world generation (Tencent)

GitHub Source ↗

DreamerV3

Reference implementation of the Dreamer family

GitHub Source ↗

TD-MPC2

Open-source TD-MPC2 codebase, 104 tasks

GitHub Source ↗

V-JEPA 2

Meta's latest JEPA world model for video understanding and robotic planning

GitHub Source ↗

I-JEPA / V-JEPA

Meta's original JEPA implementations

GitHub Source ↗

Open-Oasis

Open reproduction of Oasis Minecraft world model

GitHub Source ↗

DreamZero

Open-source WAM stack with checkpoints, eval tooling, and embodiment adaptation scripts

GitHub arXiv Project Source ↗

EnerVerse-AC

AgiBot's action-conditional embodied world model

GitHub Source ↗

MultiWorld

Scalable multi-agent multi-view video world model

GitHub Source ↗

WorldLens

WorldLens benchmark dataset + leaderboard

GitHub Source ↗

DIAMOND

Diffusion-based Atari world model + RL agent

GitHub Source ↗

LingBot-World

Open-source general world simulator with real-time interactivity

GitHub arXiv Source ↗

Micro-World

AMD open-source interactive world model for game-like environments

Project Source ↗

LeJEPA

Lean, provable JEPA self-supervised training framework (SIGReg); ~50-line core

GitHub arXiv Source ↗

Leaderboards & Benchmark Hubs

WorldLens Leaderboard

Full-spectrum driving world model evaluation

HuggingFace GitHub Source ↗

WorldArena Leaderboard

Official embodied world model leaderboard

HuggingFace Project GitHub Source ↗

WorldBench Dataset Hub

Central HuggingFace hub for videogen, occgen, and lidargen resources

HuggingFace Project Source ↗

WorldBench Dataset

Physics-centric benchmark dataset for world models and VLMs

HuggingFace Source ↗

WorldModelBench Dataset

Benchmark dataset for judging video generation models as world models

HuggingFace Project Source ↗

WBench Leaderboard

Multi-turn interactive video world model evaluation

Project GitHub Source ↗

WorldRoam-Bench Leaderboard

Long-horizon stability for interactive world models (action, vision, physics, memory). Paper: WorldOdysseyBench in Benchmarks.

Benchmarks Project Source ↗

ViewBench

Multi-view spatial-consistency benchmark for world models

arXiv Source ↗

V-JEPA 2 Model Collection

Official Meta V-JEPA 2 checkpoints (ViT-L/H/G)

HuggingFace Source ↗

WebWorld Model Collection (Qwen)

Open WebWorld-8B/14B/32B web world-model checkpoints for agent training and lookahead search

HuggingFace GitHub Source ↗

Datasets & Data Collections

EchoWM Unreal Data Pipeline

Unreal Engine pipeline separates physics trajectory collection from offline rendering to produce action-aligned multi-view video, with distributed scene screening and recovery; reports 8,767 hours across 1080p and 720p outputs.

arXiv Source ↗

Game2World / GameCleaner

Gameplay UI taxonomy, paired synthetic videos, and in-the-wild evaluation clips support temporally consistent HUD removal; a controlled pilot evaluates cleaned footage for world-model training.

arXiv GitHub Source ↗

AudioWorldSim

SoundSpaces-based open platform generates continuous binaural audio along simulated navigation rollouts, supplying reproducible acoustic training data for audio-based world models.

arXiv GitHub Source ↗

WorldArena_Robotwin2.0

Official dataset behind WorldArena embodied evaluation

HuggingFace GitHub Source ↗

EgoCS-400K

Replay-grounded egocentric Counter-Strike trajectories with video, actions, states, events, and language

arXiv Source ↗

PhysEditWorld

UE5 replay dataset for physics-editable world models with gravity interventions, actions, states, and multimodal rollouts

arXiv Project Source ↗

MobileWorld

Large-scale semantic world-model dataset for mobile GUI agents

arXiv GitHub Source ↗

MotionScape

Highly dynamic UAV-view dataset for world models

arXiv GitHub Source ↗

OmniWorld

Multi-domain, multi-modal 4D world modeling dataset and benchmark

arXiv Project GitHub HuggingFace Source ↗

EgoVerse

Large-scale egocentric human dataset for robot learning and transfer

arXiv Project Source ↗

RealWM / RealWM120K

Real-world interactive world-model data used by MagicWorld-style exploration

arXiv Source ↗

LoopNav

Spatial-consistency benchmark data for memory-aided world models

arXiv HuggingFace GitHub Source ↗

MicroVerse / MicroWorldBench

Microscale simulation data and rubric-based benchmark

arXiv GitHub Source ↗

WorldSimProbe

Public RoboTwin/LIBERO evaluation packages for action-conditioned world-model faithfulness probing

HuggingFace GitHub Source ↗

Labs, Companies & Open Stacks

Who is building what, restricted to public, primary-source material already linked in this list. "Representative entries" point to sections where the full citations and badges live. Claims about unpublished internal systems are deliberately excluded.

NVIDIA (Cosmos)

World foundation model platform for Physical AI: video WFMs, driving data engines, real-time closed-loop simulation, omnimodal Cosmos 3

Representative entries in this list: Cosmos, Cosmos-Predict2.5, Cosmos 3 (§1.6, Toolkits); Cosmos-Drive-Dreams, NVIDIA OmniDreams (§1.2.1); DreamGen, FLARE (§1.3.1); SANA-WM, FlashDreams, Causal-rCM (§1.1.3, §1.6, Toolkits)
Openness: Open weights + code for the Cosmos family

§1.6 Toolkits §1.2.1 §1.3.1 §1.1.3 Source ↗

Google DeepMind (Genie)

Foundation world models for playable environments; generalist 3D agents

Representative entries in this list: Genie, Genie 2, Genie 3 (§1.1.2); SIMA (§1.3.2); MuZero (§3.1); GraphCast (§1.5.2); Physics-IQ Verified (Benchmarks)
Openness: Papers + lab reports; Genie 2/3 are blog-documented, not open-weight

§1.1.2 §1.3.2 §3.1 §1.5.2 Benchmarks Source ↗

Meta AI (JEPA)

Non-generative predictive representation program; video JEPA world models with robot planning

Representative entries in this list: I-JEPA, V-JEPA, V-JEPA 2, V-JEPA 2.1 (§2.2); LeCun's position paper (§0.1); CWM code world model (§3.6)
Openness: Open code + checkpoints (HuggingFace collection linked in Leaderboards)

§2.2 §0.1 §3.6 Leaderboards Source ↗

Wayve (GAIA)

Generative driving world models with fine-grained controllability

Representative entries in this list: GAIA-1, GAIA-2 (§1.2.1); lab write-ups in Blogs
Openness: Papers + technical blogs; models not released

§1.2.1 Blogs Source ↗

World Labs

Spatially grounded multimodal 3D world generation

Representative entries in this list: Marble (§1.4.1); RTFM real-time frame model (Blogs)
Openness: Product + blog reports; limited technical disclosure

§1.4.1 Blogs Source ↗

1X Technologies

Humanoid robotics; world models for real-robot video prediction and policy evaluation

Representative entries in this list: 1x World Model Challenge (Benchmarks); OpenDriveLab WM Track co-listing
Openness: Public challenge + data; no full model paper in this list

Benchmarks Source ↗

AgiBot

Robotic manipulation world platforms, embodied data engines, embodied evaluation

Representative entries in this list: EnerVerse, EnerVerse-AC, Genie Envisioner, AgiBot World Colosseo (§1.3.1); EWMBench (Benchmarks)
Openness: Open code, datasets, and benchmarks

§1.3.1 Benchmarks Source ↗

Skywork AI (Matrix)

Open real-time interactive game/world stacks; 3D world generation

Representative entries in this list: Matrix-Game 1.0/2.0 (§1.1.1); Matrix-Game 3.0 (§1.1.3); Matrix-3D (§1.4.1)
Openness: Open code + weights

§1.1.1 §1.4.1 Source ↗

Tencent (Hunyuan)

Explorable, mesh-based, and simulatable 3D world generation

Representative entries in this list: HunyuanWorld 1.0, HY-World 2.0 (§1.4.1, Toolkits)
Openness: Open code + weights

§1.4.1 Toolkits Source ↗

OpenAI (Sora)

Video generation framed as world simulation — the claim that started the 2024 debate

Representative entries in this list: Video generation models as world simulators (Blogs); the debate itself: Is Sora a World Simulator?, PhyWorld (Surveys, §1.5.1)
Openness: Blog post; no open model; treated here as a position, not a system entry

Blogs Surveys §1.5.1 Source ↗

Tesla

Occupancy networks for planning, presented publicly in engineering talks (CVPR 2022 WAD keynote; no citable primary paper)

Representative entries in this list: Conceptual lineage is covered by the occupancy sections: §1.2.2, §2.3
Openness: Public talks only; nothing citable to list-standard

§1.2.2 §2.3 Source ↗

Xiaomi

Representative entries: Xiaomi EV World Model, Xiaomi-Robotics-U0, MiLA, DGGT.

Source ↗

Microsoft

Representative entries: MineWorld, Latent Spatial Memory.

Source ↗

Alibaba/Qwen and DAMO

Representative entries: WorldVLA, Qwen-RobotWorld, Qwen-AgentWorld, WorldOlympiad.

Source ↗

SenseTime

Representative entries: OpenDWM, MaskGWM, UniMLVG.

Source ↗

GigaAI

Representative entries: ReconDreamer, GigaWorld, GigaBrain.

Source ↗

Meituan

Representative entries: WBench.

Source ↗

Ant/Robbyant

Representative entries: LingBot-VA, LingBot-World.

Source ↗

Selected Technical Blogs & Reports

This section is intentionally selective. It favors official research-lab posts, technical reports, and a small number of high-signal explainers over general-audience trend pieces.