In the development of UAV systems, complex environments, high trial and error costs, and long joint debugging cycles are common pain points for almost all teams.

This upgrade of PrometheusSim (ProSim) will Pod visual landing and End-to-end obstacle avoidance Installed into simulation, combined with Windows one-click deployment, UE4 high-fidelity scenes, multi-sensors and formation/task links, ProSim supports a complete closed loop from algorithm development, simulation verification to real machine migration. It helps you move more risks and time costs to the screen and try to only do the "last mile" verification on the real machine.

The ProSim installation package and wiki instructions are now available. See the end of the article for how to obtain them. At the same time, you are welcome to join the ProSim learning and exchange group to exchange practical experiences with developers.

Pod visual landing

• Industrial grade precise localization: Through the collaborative control of the three-axis gimbal (pitch/roll/yaw), the UAV can dynamically adjust its attitude to maintain target lock.

• Sub-meter landing accuracy: Combining the visual recognition and flight controller fusion algorithm, the landing localization error is less than 5cm.

• One-click trigger operation: Clicking on the target through the video stream can automatically complete the entire process of identification, tracking, and landing.

• Lightweight solution: A low-cost solution that only requires yaw axis control.

• One-click trigger operation: Click on the target through the ground control station to automatically complete the entire process of identification, tracking, and landing.

• Yaw tracking: Through visual estimation, target yaw tracking is achieved.

Advantages:

1. The pod system uses Prometheus visual-inertial joint localization, which can maintain heading memory even if the target is temporarily blocked.

2. There is no requirement for the size of the landing board, and no calibration is required for visual target recognition estimation.。

3. Optimize the control algorithm to adapt to target recognition jitter scenarios.

4. Integrate estimated target position, three-axis/single-axis pod control, and landing guidance.

5. The visual tracking and landing logic is more suitable for actual engineering scenarios.

End-to-end obstacle avoidance

Bionic visual decision-making model

• Human-like decision-making mechanism: based on YOPO(You Only Plan Once) algorithm realizes end-to-end obstacle avoidance of "awareness to decision-making".

• Dynamic scene adaptation: It has a high success rate of obstacle avoidance in complex environments such as unmapped woods/construction sites.

• Millisecond response: Model inference latency is <3ms.

Full-process simulation training system

• Multimodal acquisition:

Construct diverse obstacle scenes (forests/buildings, etc.) in PrometheusSim to simultaneously collect global radar point cloud + depth image data.

• Scale data:

Generate tens of thousands of labeled data sets: 15,000+ scene samples, covering different weather conditions.

Extreme performance inference engine

• TensorRT accelerates deployment:

1. The Prometheus warehouse provides one-click conversion scripts:

yopo_trt_transfer.py

2. Support FP16 quantification: the model volume is reduced by 50%, and the inference speed is increased by 3 times.

Out-of-the-box model library

• Pre-conversion model path:

/home/amov/Prometheus/Modules/YOPO/run/yopo_trt.pth

Seamless migration to real devices

• The simulation verification model can directly deploy the physical UAV

Continuously expanding simulation ecosystem

ProSim has built simulation capabilities covering the entire cycle of UAV R&D, based on modularity and unified interfaces, to facilitate continuous access to new scenarios, sensors and algorithms in the same simulation domain. In addition to this expanded Pod visual landing and YOPO end-to-end obstacle avoidance In addition to the functions, it also supports the following core functions:

Cluster collaboration

Intelligent formation system: leader-follower architecture; default formations include straight and triangular formations

Group obstacle avoidance algorithm: supported Multi-camera depth camera and Multi-aircraft radar point cloud Joint obstacle avoidance planning

environmental awareness

Multi-sensor planning:

• Livox Mid-360LiDAR obstacle avoidancemapping

• Intel D435i depth camera obstacle avoidance

• 2D LiDAR obstacle avoidance

Dynamic target tracking:

• YOLOv5 human body recognition tracking

core flight controls

Multi-mode takeoff and landing control:

• One-click takeoff/landing/return

• Precise hovering (localization accuracy ±0.1m)

Full attitude flight:

• Body coordinate system control (front/back/left/right/lift)

• Inertial coordinate system control (global XYZ position)

• Longitude and latitude height control (GPS localization mode)

Autonomous trajectory tracking:

• Default trajectory: circle, figure 8

• Waypoint mission: compatible with Prometheus ground control station Visual planning

Resource Express

Follow the official account and send a private message to ProSim in the background to obtain the installation package and wiki link.

Previous ProSim introduction articles:

Get started with ProSim’s new features in half an hour, and play new tricks in UAV simulation!

PrometheusSim simulation platform is online! | Build a UE4 super-realistic test flight ground in seconds