Decart AI Lab

Decart AI Lab builds real-time world models and optimization infrastructure for Physical AI, including Oasis 3 for interactive simulation and DOS for low-latency AI workloads.

Andere

Website Besuchen

Overview

Decart AI Lab builds real-time world models and optimization infrastructure for generative AI systems. Its site centers on three product lines: Oasis for physical AI world generation, Lucy for real-time world editing, and DOS, the Decart Optimization Stack for low-latency inference and training.

Oasis 3 is the clearest productized world model on the site. It generates realistic, controllable simulation environments in real time, responds to physical control signals, and is positioned for robotics and autonomous vehicle workflows that need repeated, closed-loop training in many scenarios.

Core capabilities

Real-time world generation

Generates realistic, physically grounded environments that evolve continuously instead of producing a one-off offline clip.

Action-conditioned feedback

Supports closed-loop simulation by reacting to steering, movement, and API commands as a policy acts inside the environment.

Multi-view output

Produces synchronized multi-camera views from one coherent world state, helping maintain consistency across angles.

API access

Exposes the world model through an API, so teams can connect it to existing simulation and training pipelines.

Low-latency response

Runs with low latency that Decart cites at under 200 ms end-to-end, making interactive training loops practical.

Hardware-aware infrastructure

Covers optimization across GPUs, TPUs, Trainium, AMD, and other accelerators through the Decart Optimization Stack.

Where it fits

Autonomous vehicle training
Train autonomous driving policies inside prompt-driven environments that can vary geography, weather, lighting, traffic, and driving scenarios without needing to record new footage each time.
Robotics simulation
Test robots in closed-loop simulation where actions immediately affect the next world state, supporting repeated practice and online reinforcement learning.
Physical AI scenario generation
Build and evaluate scenarios for drones, off-road vehicles, maritime systems, or humanoid manipulation when real-world edge cases are hard to capture safely.
Low-latency AI optimization
Use the Decart Optimization Stack to tune latency, throughput, and cost across GPU, TPU, Trainium, and AMD-based workloads.
Developer pipeline integration
Integrate the model into existing simulation workflows through API access and SDK-based setup.

Pros and Cons

Pros

Real-time, interactive simulation with continuous world state rather than static generated video.
API access makes it usable inside existing robotics and training pipelines.
Multi-view output and action-conditioned responses support closed-loop testing.
The site clearly states a first target use case in autonomous vehicles and a broader path to other physical AI domains.
DOS is designed to optimize latency, throughput, and cost across a wide hardware stack.

Cons

The pricing page currently returns a 404, so pricing and access terms are not published there.
The source emphasizes physical AI and simulation; it does not show broad support for unrelated general-purpose product categories.
Some product details are still framed as research or launch announcements rather than full documentation, so implementation specifics are limited on the public pages.

FAQ

What is Oasis 3?

Oasis 3 is an interactive world model for Physical AI. It generates realistic, controllable, multi-view simulation environments in real time and responds to physical control signals for closed-loop training.

How do teams use Oasis 3?

Decart describes Oasis 3 as available via API, with SDK access shown in the quick-start example on the product page. The page also presents it as infrastructure for developers, researchers, and robotics teams.

Which applications does Oasis 3 target first?

The page says Oasis 3 is designed for autonomous vehicles first, with the same system extending to drones, off-road vehicles, maritime applications, and humanoid manipulation.

How fast is Oasis 3?

The product page states under 200 ms end-to-end latency and 22 FPS at 512×768×3, while also noting that Decart continues optimizing the stack.

What is the Decart Optimization Stack?

Decart’s optimization stack is a vertically integrated platform for LLM, agentic, video, and world model workloads. It spans hardware-aware model design, kernel tooling, proprietary compilers, and inference optimization, and integrates into existing infrastructure.

Quick Facts

Category: AI infrastructure / world models
Primary product: Oasis 3
Related products: Lucy 2.0 and the Decart Optimization Stack (DOS)
Access model: API-accessible
Primary users: Robotics teams, AV companies, researchers, and developers
Source domain: decart.ai