World Models (Self-Supervised Video, Robotics)
Predictive representation models for planning.
Tier Grades
Tier A: Frontier AI labs
feasible
high confidence
Meta and NVIDIA are observed and released open weights.
Tier B: Indonesian AI labs
stretch
medium confidence
Pretraining on 1M hours of video is out of reach on the evidence, but the action-conditioned stage needs only 62 hours of robot data on open weights.
Tier C: Institution units such as ITB
stretch
medium confidence
The 62-hour robot-data stage on an open 1.2B model is small, but robotics hardware and compute are not given in the notes.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
stretch
low confidence
The low data need on open weights suggests a possible proof of technology, but the notes give no compute or hardware anchor.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeV-JEPA 2: 1.2B parameters; Cosmos sizes and compute not stated.
- DataV-JEPA 2: over 1M hours of video plus 1M images, then 62 hours of robot data.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): not sourced in the notes.
- Time horizonMulti-year.
- Shelf-life signalsnot in the notes
- Pipeline roleTier A releases open world models; Tier B and Tier C adapt them with small robot datasets; Tier D could carry them into robotics proofs of technology.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
V-JEPA 2NVIDIA CosmosObserved actors
Linked models
V-JEPA 2world modelexemplar
Cosmosworld modelexemplar
Gemini Robotics-ER 2multimodal LLMused by
GigaWorld-1world modelproduced byCosmos 3world modelproduced byCosmos 3 Nanoworld modelproduced byCosmos 3 Superworld modelproduced byGemini Robotics-ER 1.6multimodal LLMused byCosmos Predict 2.5world modelproduced byCosmos Transfer 2.5world modelproduced byCosmos Reason 2multimodal LLMused by
GWM-1 Roboticsworld modelproduced byGemini Robotics-ER 1.5multimodal LLMused byV-JEPA 2-ACrobot policyproduced byGemini Robotics-ERmultimodal LLMused byV-JEPAworld modelproduced byCosmos 3 Edgeworld modelproduced byCosmos Predict 1world modelproduced byCosmos Predict 2world modelproduced byCosmos Transfer 1world modelproduced by
UnifoLM-WMA-0world modelproduced byCosmos Reason 1multimodal LLMused byEagle 2.5multimodal LLMused byLinks
R&D TopicWorld Models (Self-Supervised Video, Robotics)
ModelsV-JEPA 2Meta, world modelCosmosNVIDIA, world modelV-JEPA 2-ACMeta, robot policyV-JEPAMeta, world model+19 more
V-JEPA 2Meta, world modelCosmosNVIDIA, world modelV-JEPA 2-ACMeta, robot policyV-JEPAMeta, world model+19 moreWho Runs Itnone recorded
Models 23
CosmosNVIDIA, world modelexemplarGemini Robotics-ER 1.6Google DeepMind, Boston Dynamics, multimodal LLMused byCosmos 3NVIDIA, world modelproduced byCosmos Reason 2NVIDIA, multimodal LLMused byCosmos Predict 2NVIDIA, world modelproduced byCosmos Predict 2.5NVIDIA, world modelproduced byCosmos Reason 1NVIDIA, multimodal LLMused byGemini Robotics-ERGoogle DeepMind, multimodal LLMused byGemini Robotics-ER 1.5Google DeepMind, multimodal LLMused byGemini Robotics-ER 2Google DeepMind, multimodal LLMused byCosmos Predict 1NVIDIA, world modelproduced byCosmos Transfer 1NVIDIA, world modelproduced by+11
Cosmos Transfer 2.5NVIDIA, world modelproduced byEagle 2.5NVIDIA, multimodal LLMused byGWM-1 RoboticsRunway, world modelproduced byV-JEPAMeta, world modelproduced byV-JEPA 2Meta, world modelexemplarV-JEPA 2-ACMeta, robot policyproduced byCosmos 3 EdgeNVIDIA, world modelproduced byCosmos 3 NanoNVIDIA, world modelproduced byCosmos 3 SuperNVIDIA, world modelproduced byGigaWorld-1GigaAI, Tsinghua University, world modelproduced byUnifoLM-WMA-0Unitree Robotics, world modelproduced byPapers
- Self-Supervised Learning from Images with a Joint-Embedding Predictive ArchitectureMahmoud Assran et al., arXiv preprint, 2023 founding
- V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and PlanningMido Assran et al., arXiv preprint, 2025 state of the art
- Revisiting Feature Prediction for Learning Visual Representations from VideoAdrien Bardes et al., arXiv preprint, 2024 state of the art
- Mastering Diverse Domains through World ModelsDanijar Hafner et al., arXiv preprint, 2023 state of the art
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- Introducing the V-JEPA 2 world model and new benchmarks for physical reasoning ai.meta.com https://ai.meta.com/blog/v-jepa-2-world-model-benchmarks/
- Cosmos World Foundation Model Platform for Physical AI arxiv.org https://arxiv.org/abs/2501.03575
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "World Models (Self-Supervised Video, Robotics)", https://aurion.semai.id/topics/world-models-self-supervised-video-robotics/, records of October 2026.