The Interactive World Model for Physical AI
Oasis 3 generates realistic, controllable, multi-view simulation environments in real-time, where robots can learn, act and improve across infinite scenarios, including the edge cases that matter most. Starting with autonomous vehicles.
We train for every world. Before it matters.
The first interactive world model available via API. Realistic, physically grounded environments where autonomous systems can train, test, fail, and improve safely, responding to physical control signals, not just keyboard input.
Any scenario, available on demand in environments that run and evolve continuously.
Multiple viewpoints generated from one coherent world state, consistent across every angle.
The only interactive world model available for teams to build with from day 0.
Under 200ms end-to-end. Low latency means the world responds fast enough for real training loops.
The real world doesn’t give second takes
Real-world training is slow, expensive, and hard to scale. The scenarios that matter most, edge cases, rare failures, dangerous conditions are exactly the ones that are hardest to capture.
Oasis 3 gives you every scenario, on demand. Because the best preparation is infinite practice.
How it works
Prompt a world, train inside it and never stop learning
Prompt the environment
Define the world: geography, weather, lighting, traffic, and the scenarios your system needs to handle.
Send control signals
Let the agent act. Oasis 3 responds to steering, movement, and API commands in real time, creating a true closed-loop simulation.
Train continuously
Run scenarios thousands of times or change conditions on the fly without the need to record or reset.
Autonomous vehicles first. Every robot that comes after.
Oasis 3’s first application focuses on autonomous vehicles: prompt-driven environments across different geographies, weather conditions, and driving scenarios, navigable in real time.
The same system extends across physical AI. Drones, off-road vehicles, maritime, humanoid manipulation anywhere the real world can't provide the scenarios robots need.
World models, now accessible as infrastructure
Oasis 3 launches as infrastructure, not just a research demo — giving developers, researchers, and robotics teams direct access to real-time world generation.
import importlib, os, subprocess, sys
# The SDK (`decart-oasis`) is installed from the public PyPI package. `oasis_demo` — the extras you
# can play with (reward, env, preview) — isn't published, so we clone it from GitHub and install it
# editable, so you can tweak e.g. `extras/oasis_demo/rl/reward.py`, restart the runtime, and re-run.
REPO_URL = "https://github.com/DecartAI/decart-robotics.git"
REPO_DIR = REPO_URL.rstrip("/").split("/")[-1].removesuffix(".git")
if importlib.util.find_spec("oasis_demo") is None:
subprocess.run([sys.executable, "-m", "pip", "install", "-q", "decart-oasis"], check=True)
if not os.path.isdir(REPO_DIR):
subprocess.run(["git", "clone", "--depth", "1", REPO_URL], check=True)
subprocess.run([sys.executable, "-m", "pip", "install", "-q", "-e", f"{REPO_DIR}/extras"],
check=True)
# Editable installs register via a .pth that's only read at interpreter startup, so put the
# source on sys.path to import it now without restarting the runtime.
sys.path.insert(0, os.path.abspath(f"{REPO_DIR}/extras"))
importlib.invalidate_caches()
# API key (required to connect). Set it in the Colab secrets panel (🔑) as DECART_API_KEY,
# or here with os.environ["DECART_API_KEY"] = "...".
if not os.environ.get("DECART_API_KEY"): try:
from google.colab import userdata
os.environ["DECART_API_KEY"] = userdata.get("DECART_API_KEY") or ""
except Exception:
pass
print("DECART_API_KEY set:", bool(os.environ.get("DECART_API_KEY")))
from decart_oasis import A2VClient
from oasis_demo.live_preview import LiveCameraPreview
print("Ready.")
Built on the stack that makes real-time possible
Oasis 3 runs on DOS, Decart’s ultra-optimized inference and training platform that makes real-time world generation economically viable at scale. The same stack powers Decart's own models and is in production with the world's largest hyperscalers and chip manufacturers.
Build physical AI that’s ready for any scenario.
We’re partnering with robotics teams, AV companies, and research groups to shape the next generation of interactive world models.