Decart Launches Oasis 3: A Realistic, Low‑Latency World Model for Robotics Training

Frontier AI research lab Decart has introduced Oasis 3, a world model designed to close the divide between synthetic simulation and real‑world AI deployment.

The video‑output engine, unveiled this month, accelerates training of robot and autonomous‑vehicle control systems, equipping them to thrive amid unpredictable real‑world conditions.

Robotics developers confront a scarcity of high‑quality data necessary to train systems that can navigate complex, real‑world environments.

While a vehicle can learn to maneuver a static parking lot with fixed traffic cones, the open road presents a far more demanding setting where weather, lighting, and dynamic obstacles vary constantly.

Training systems to handle the chaos of urban streets—heavy rain, sudden obstacles, erratic traffic—represents a distinct challenge that Oasis 3 tackles.

The robotics training bottleneck

Large language models (LLMs) have surged ahead, but general‑purpose robotics—or physical AI—has lagged, largely due to a shortage of rich media resources.

Bessemer Ventures noted that LLM developers benefit from scraping billions of public web pages, a luxury unavailable to Vision‑Language‑Action (VLA) models that must interpret and act within physical spaces.

VLA models ingest environmental data, process it, and then respond. Training them offers three primary paths:

• Teleoperation—human operators mimic robot actions in a controlled suit. While yielding the highest‑quality data, it is prohibitively expensive and slow, making large‑scale deployment impractical.
• Open‑web videos—readily available but messy, lacking consistent environments, spatial telemetry, and direct action conditioning.
• Synthetic data—a middle ground, yet current physics engines fall short of real‑world nuance, leading to the so‑called sim‑to‑real gap.

That gap manifests when real‑world randomness—oil spills, fragile packaging, unexpected debris—throws autonomous systems off‑balance, revealing their limitations.

Closing the gap with closed‑loop, generative simulations

Decart claims Oasis 3 bridges existing virtual training limits by fusing photorealistic motion graphics with a robust physics engine.

Embedded in a single, high‑performance training loop, Oasis 3 produces action‑conditioned video streams that can generate virtually any chaotic scenario developers envision, creating a training environment that closely mirrors reality.

The platform supports multiview, ultra‑realistic environments that are fully controllable; a self‑driving car’s lateral deviation triggers a generative stream that adjusts perspectives in under 200 ms—well within reinforcement learning requirements.

Co‑designed with Nvidia’s physical‑AI ecosystem, Oasis 3 runs on CoreWeave’s specialized cloud infrastructure at 22 fps, delivering interactive virtual environments at 512×768×3 resolution.

It offers a native three‑camera view to preserve spatial and temporal consistency from multiple angles, enabling autonomous systems to gauge depth and peripheral context accurately.

Oasis 3 is accessible via Decart’s API, allowing developers to integrate it seamlessly into their existing physical‑AI simulation pipelines.

Training robots to conquer uncharted territory

Achieving science‑fiction‑level humanoids requires training robots to manage unique edge cases in real time—situations impossible to replicate in a laboratory, such as a load falling onto a road while an autonomous vehicle’s camera is obscured by mud.

Oasis 3 empowers developers to create infinite variations of such events using simple natural‑language prompts, spanning different angles, weather conditions, and road surfaces.

By exposing models to millions of hazards affordably, developers can ensure readiness for any plausible real‑world scenario.