Physical AI / embodied world models
Decart Oasis 3 Preview
Oasis 3 Preview is Decart's real-time, promptable world model you drive: set a scene with text, stream throttle/steering actions, and it generates the next camera frames (left / front / right). Because it turns actions into frames in real time, it doubles as a learned driving simulator you can run reinforcement learning inside.
Overview
| Status | Tracked |
|---|---|
| Access | API available |
| Released | 2026 |
| Inputs | text prompt (scene), throttle + steering actions |
| Outputs | multi-camera RGB frames (left / front / right) |
| Best for | driving-sim & autonomy research, reinforcement-learning environments, embodied-AI experiments, action-conditioned world modeling, robotics prototyping |
Why it matters
It points real-time world models at robotics and embodied AI rather than entertainment — a hosted, promptable driving simulator you can train policies in without owning GPUs.
Roamscape use
Tracked in Roamscape's model hub as a robotics/RL-focused world model. Not wired into the browser /live workspace — its Python-gRPC + driving-action interface fits agents and simulators, not WebRTC exploration.
Strengths
- real-time action→frame generation
- promptable scenes
- multi-camera output
- ready-made RL path (Gymnasium / SB3)
- hosted — no model to run
Limitations
- Python gRPC SDK only (no browser/JS real-time SDK)
- driving actions (throttle/steering), not free camera exploration
- priced comparably to premium real-time models
- preview stage