uav-world-model
Imagination-based UAV world model (LeWM, JEPA+SIGreg): imagines future danger from drone-POV video so a CEM planner proactively avoids turret threats before single-frame detection can react. Isaac Sim + PyTorch.
Overview
PLAN — Predictive Latent Autonomous Navigation.
Imagination beats detection for drone danger. A world model that rolls its latent forward sees a threat a single-frame detector can't, and acting on that imagination keeps a drone alive.
Sim-only PoC. Synthetic urban environment, NVIDIA Isaac Sim 2.3.2, 2× RTX A6000. Trained weights are not redistributed; produce them from source. Apache-2.0.
A turret that is visible but not yet aimed is not yet dangerous. A turret about to aim is dangerous before any single frame shows it. A detector reads one frame and asks whether there is a threat now based on the latent representation of that frame. A world model thinks ahead in latent space, and asks whether it will be in danger in the future if it follows it's current planned course.
Clone with the swm submodule (LeWM, SIGreg, stock CEM planner):
Open the showcase in Foxglove (Mac, no GPU):
From the project’s README on GitHub.
At a glance
| Repository | Kentucky-Open-Science/uav-world-model |
|---|---|
| Research area | Robotics, drones & sensing |
| Primary language | Python |
| Languages | Python 91.7%, Shell 8.3% |
| License | Apache-2.0 |
| Stars / forks | 0 / 0 |
| Open issues and pull requests | 0 |
| Created | 2026-07-21 |
| Last push | 2026-08-26 |
| Default branch | main |
| Topics | drone, isaac-sim, jepa, object-detection, pytorch, robotics, self-supervised-learning, world-model, world-modeling, world-models |
What the README covers
- Situation
- Result 1 — the world model predicts ahead of detection
- Result 2 — acting on imagination saves you
- How it works
- Data collection and training
- How to run
- Limitations
- Stack
Top contributors
- @EvanDamron (7 commits)
Related projects
-
Temi-VOC-Datasets
VOC/smell data from ~April 2025 to September 2025 recorded by Robot Temi
-
HAVOC-TemiApp
HAVOC is an advanced Android application designed to transform the Temi robot into an autonomous healthcare assistant. The project leverages Large Language Models (LLMs), …
-
AirSimVideo
Archived
Manipulating an AirSim simulation remotely and extracting video
-
TrafficStream2026
Lex Traffic + YOLO
-
Yolo-Person-Detection
This repository contains multiple Python scripts utilizing the YOLO model for various real-time object detection and pose estimation tasks, including fall detection and general …
-
yolo_lesion
Archived
No description on GitHub yet.
Get the code
git clone https://github.com/Kentucky-Open-Science/uav-world-model.git