Decart’s new world model can simulate hours of photorealistic driving — with some caveats

2026-06-10 · Source: AI News & Artificial Intelligence | TechCrunch · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Emerging Technologies & Innovation · Depth: Intermediate, medium

Summary

Decart, an AI startup, has launched Oasis 3, an interactive world model available via API that generates photorealistic driving environments in real time. Priced at \$0.02 per second, Oasis 3 is initially aimed at autonomous vehicle companies for simulating rare scenarios at scale, with plans to expand into robotics. The model builds on Decart's Lucy real-time video model and is designed to foster a developer ecosystem, similar to OpenAI's approach with language models. Decart recently raised \$300 million, boosting its valuation to nearly \$4 billion, with strategic investors including Toyota, Adobe, and eBay. Oasis 3 distinguishes itself with photorealism and infinite generation capabilities, powered by Decart's DOS software, which enables efficient operation on Nvidia, Amazon, and Google hardware, making it significantly cheaper to run than competitors. However, the model currently faces limitations, including thematic integrity degradation over extended simulations, unresponsive controls, and a lack of proper physics simulation (e.g., cars driving through objects). These issues stem from its auto-regressive architecture and limited memory context, which Decart is actively working to improve.

Key takeaway

For autonomous vehicle engineers evaluating simulation platforms, Decart's Oasis 3 offers a cost-effective API for generating photorealistic driving environments infinitely. Its efficiency and real-time capabilities are valuable for initial scenario generation and testing diverse edge cases. However, its current auto-regressive architecture struggles with long-term thematic consistency and physics accuracy. Consider integrating Oasis 3 for rapid prototyping, but plan for post-processing or alternative solutions to address physics and consistency gaps.

Key insights

Decart's Oasis 3 offers real-time, photorealistic world simulation via API, targeting autonomous vehicle development despite current physics and consistency limitations.

Principles

• Vertical integration optimizes model efficiency and cost.
• Auto-regressive generation impacts long-term consistency.
• Developer ecosystems drive model adoption and innovation.

Method

Oasis 3 generates physically accurate, multi-camera environments frame-by-frame using an auto-regressive architecture, looking back at previous frames to decide next steps.

In practice

• Simulate rare driving scenarios for AV training.
• Develop physical AI applications via API access.
• Test edge cases with infinite environment generation.

Topics

• World Models
• Autonomous Vehicles
• Real-time Simulation
• API Economy
• Physical AI
• Decart Oasis 3

Best for: Investor, CTO, VP of Engineering/Data, AI Engineer, Robotics Engineer, Director of AI/ML

Related on AIssential

• Decart lays the foundation for physical AI systems with Oasis 3 — Oasis 3 provides hyper-realistic, real-time simulation environments critical for training autonomous physical AI systems.
• NVIDIA OmniDreams: Real-Time Generative World Model for Closed-Loop Autonomous Vehicle Simulation — OmniDreams is a generative world model for AV simulation that overcomes data limitations and can serve as a policy architecture backbone.
• World Models in Plain English — World models aim to spatially compress and predict real-world environments, moving beyond text-based learning.
• NVIDIA Enables the Next Era Of Physical AI Research With Agent Skills For Autonomous Vehicles, Robotics And Vision AI — Physical AI development requires comprehensive workflows beyond just model strength, integrating scene reconstruction and edge-case generation.
• Causality-Aware End-to-End Autonomous Driving via Ego-Centric Joint Scene Modeling — Causality-aware modeling improves end-to-end autonomous driving by capturing ego-agent interdependencies for consistent trajectory prediction.
• RoboDream: Compositional World Models for Scalable Robot Data Synthesis — A generalizable embodiment-centric world model synthesizes photorealistic robot demonstrations by decoupling motion from environment synthesis.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI News & Artificial Intelligence | TechCrunch.