Summary
Decart, a two-year-old startup recently valued at nearly 300M raise backed by Toyota, Adobe, Nvidia, Sequoia, and Benchmark (among others), launched Oasis 3, an interactive world model that generates photorealistic, multi-camera driving environments in real time and is offered via API at roughly $0.02/second. It targets autonomous-vehicle companies wanting to simulate rare or dangerous edge-case scenarios cheaply, with an explicit roadmap toward robotics and other physical-AI use cases.
Key Contributions
- A real-time, interactive world model capable of generating extended (multi-hour) photorealistic driving sessions from a single prompt, rather than short clips.
- Multi-camera, “infinite” generation designed specifically for simulating rare/dangerous AV edge cases without real-world risk.
- Decart’s proprietary “DOS” (Decart Optimization Stack), vertically tuned across Nvidia/Amazon/Google hardware, claimed to cut inference cost by more than an order of magnitude versus competitors.
- Commercial API productization ($0.02/s pricing) aimed squarely at AV developers as a first market, with robotics/physical AI as a stated expansion target.
Strengths
- Reported to produce the most photorealistic single-prompt environments relative to comparable interactive world models (e.g., Google’s Genie 3, World Labs’s Marble), per TechCrunch’s hands-on comparison.
- Real-time interactivity plus multi-hour session length is a meaningfully different capability class than prior short-clip video world models, directly useful for long-tail scenario coverage in AV testing.
- Strong strategic backing (Toyota, Adobe Ventures, Toyota Ventures, Nvidia, Sequoia, Benchmark) signals real industry interest in this as a simulation-layer play, not just a research demo.
- Cost-optimized inference stack targets the actual bottleneck for this class of product (compute cost per simulated second), which is the right lever if the goal is to make world-model-based simulation commercially viable at scale.
Weaknesses
- TechCrunch’s hands-on testing found significant degradation over extended sessions: thematic/environmental consistency breaks down as the user moves through the world (a prompted “New York City street” drifted into a generic Western city), and returning to an earlier location does not reproduce it — the world is not persistent or spatially consistent, it is regenerated locally.
- Physics is not properly simulated: testers observed cars passing through objects, which is a serious problem for a product explicitly marketed for safety-critical AV edge-case testing.
- Controls were reported as unresponsive/laggy, undermining the “interactive” framing for real-time decision-relevant simulation.
- These are precisely the properties (spatial consistency, physical plausibility, controllability) that matter most for the stated use case (testing rare/dangerous driving scenarios) — a photorealistic but physically incoherent simulator risks producing edge cases that are visually convincing but dynamically meaningless, which could be worse than no simulation if used naively for AV validation.
Open Questions
- Are AV customers actually using Oasis 3 for anything beyond visual/perception-model stress testing, given the acknowledged lack of physical grounding — or is closed-loop planner/control testing still out of reach?
- How does Decart plan to address world persistence and physical consistency, and is that even tractable within a purely generative (non-physics-engine) video-world-model architecture, or does it require hybridizing with a physics simulator?
- What does the promised pivot to robotics/physical-AI look like concretely, given manipulation tasks are typically far less forgiving of physics violations than driving corridor scenarios?
- No independent, quantitative benchmark numbers are cited in press coverage — how does Oasis 3 compare to Genie 3 / Marble / Cosmos on standard world-model benchmarks (e.g., WorldModelBench) rather than qualitative impressions?
Significance
Oasis 3 is a visible commercial bet that photorealistic, real-time interactive world models can become a paid simulation layer for physical AI (starting with autonomous driving), and the TechCrunch review is a useful public data point on how far current generative world models still are from physically consistent, persistent simulation — a gap that is central to whether such models can be trusted for safety-relevant testing rather than just visual demos.