TJU-Aerial-Robotics/YOPO

What it solves

YOPO addresses the challenge of autonomous drone navigation in obstacle-dense environments. It replaces the traditional multi-step pipeline—which typically separates perception, path searching, and trajectory optimization—with a single, unified neural network to reduce computational overhead and avoid local minima during planning.

How it works

The system functions as a "one-stage" planner, drawing inspiration from object detectors like YOLO. It uses a set of motion primitives as anchors to cover the potential search space and then predicts the necessary offsets and scores for these primitives to refine the trajectory.

Unlike imitation learning (which requires expert data) or reinforcement learning (which requires trial-and-error), YOPO uses a "guidance learning" strategy. It directly back-propagates the gradients of trajectory costs (such as those from Euclidean Signed Distance Fields or ESDF) into the network weights to optimize the path.

Who it’s for

This project is designed for robotics researchers and engineers working on agile drone flight, autonomous navigation, and sim-to-real transfer for aerial vehicles.

Highlights

• Unified Architecture: Integrates perception, mapping, and optimization into one network.
• Efficient Training: Uses guidance learning to avoid the need for expensive simulators or expert demonstrations.
• High Performance: Supports TensorRT deployment for ultra-fast inference (1-5ms on NVIDIA Orin NX).
• Sim-to-Real Capability: Policies trained on ground truth depth images can be zero-shot transferred to real-world stereo cameras and unseen environments.