This year's SUAS mission was very much alike the previous year's, with some essential modifications that made it far more practical to take what we've learned from Fenrir and make a new drone. The new drone is fittingly named Skøll, after one of the sons of the wolf god Fenrir in Norse mythology.

Though we kept a lot of the general ideas of Fenrir on Skøll, a few significant changes were made. In theme with Skøll being Fenrir's son, the new drone is smaller and weighs significantly less, as the updated mission rules no longer require us to do everything in a single run, opening up the possibility for a pit-stop during which we swap the main batteries giving power to the motors. The most notable change in construction is that Skøll has overlapping propellers, a choice made to follow the newest developments and cutting-edge research in the world of drones, making the design more compact while increasing stability and energy efficiency.

The radars are no longer a part of the sensor payload, because a RemoteID module was made a requirement as part of the competition, and one such module was mounted on Skøll. These RemoteID modules constantly send information about the drone they are mounted on to anyone and anything that picks up the stream. Amongst other things, this information consists of the drone's position and speed, which are variables we are able to use in a completely new iteration of the avoidance algorithm, based on efficient path-finding algorithms.

Another change in the construction is that the downward facing camera is now mounted on a gimbal that we developed to stabilize the camera even in very fast flight. This ensures that we get clear and sharp images for our perception pipelines that also underwent a significant upgrade for more reliable performance. A big point we constantly kept in our minds when making Skøll was ease of handling, so we ensured that all of the systems are readily available for swapping, repair, and that they are easy to use. This is most reflected in the updated designs of the drop off system and the design of the battery swap system, made to minimize the time spent on the ground.

Fenrir was our first attempt at completing the mission given at the SUAS competition.

With its almost 25kg of maximum take-off weight, it is by far the biggest drone we have built yet. Fenrir was our first deviation from the traditional quad and octo-quad configurations; we made it a hexacopter to ensure high payload capacity and redundancy should a motor give way in the air. It was built to fly an incredible 20km in a single run between 20min and 30min. After having successfully flown the great distance, Fenrir had the task of scanning a large area looking for specific markers that denoted where payload was supposed to be released.

The payload was in the form of 5 half-litre water-bottles for which we designed and made our own drop-off winch-based system to ensure precise and safe release. The drop points were determined using a high-resolution camera from Allied Vision whose camera feed was fed to a custom-made neural network object detection and classification system. These detections were then combined with flight controller GPS data in a localisation pipeline to precisely pinpoint the targets. Just like with all our other drones, the data processing on Fenrir was done onboard, this time upgrading to the state-of-the-art Orin NX computer from NVIDIA.

What also made Fenrir special was that SUAS was the first competition to require actively avoiding other competitiors' drones while performing the mission, so we made our very own radar-based drone avoidance system. Fenrir had 7 radars mounted all around to ensure coverage from all sides, and the algorithm was made in such a way to always ensure a safe distance from the other drones flying about in the competition area.

Solan & Ludvig is the first UAV duo made for the TAC challenge 2022.

Solan and Ludvig are equipped with different sensors for two scenarios. First, Solan comes with a Zed stereo camera attached to the mainframe; The camera utilizes 3D sensing technology to help Solan inspect anomalies from the top of an industrial site. On the contrary, Ludvig carries a Livox Avia LiDAR scanner to assemble 3D models for terrain mappings.

That said, both Solan & Ludvig share the same configurations and can carry out each other’s mission by replacing the corresponding sensor. Thus, Solan & Ludvig are almost identical; the only difference is the sensor payload. In 2022, the new TAC challenge required the drone to adjust the sensor payload without being overweight. Therefore, vanilla Ludvig is heavier due to the installment of a Lidar compared to a lighter camera on Solan. Still, both drones are within the payload limit. Furthermore, Solan & Ludvig can autonomously resume flight from the last waypoint during a battery swap compared to Marlin.

Marlin is our first outdoor drone designed for remote operations for TAC challenges.

Adapting to TAC mission scenarios, Marlin is our first drone to open up a new software pipeline in autonomous operations: AI decision-making and autopilot. The built-in AI system utilizes a “Superfluid” state machine and MAVROS package to enable extended flight beyond visual line of sight (BVLOS). Marlin employs thermal imagining to detect heat signatures from human-like objects for search and rescue scenarios. Other key characteristics include a powerline avoidance system and a failsafe mechanism.

The main frame is equipped with a custom antenna and various sensors for a gimbal, RGB thermal camera, Real Time Kinematic (RTK), and Pixhawk-based flight controller. The key takeaway from the name Marlin is its survivability under harsh weather conditions. This means that external hardware is made of waterproofing materials. In addition, Marlin is a lightweight A3 drone weighing only 2.4 kg with all essential accessories. Adapting to TAC mission scenarios, the drone can carry out regular flights with an additional 2.2 kg weight.

After winning first prize at the 2021 TAC challenge, Marlin is now retired as an outdoor test drone.

‍

Nostromo is Ascend’s longest-established drone for IARC Mission 9 from 2020 to 2022. In 2020, the development was extended for a year due to COVID restrictions. This granted us time for massive upgrades from the original prototype.

Hardware accessibility is comparably faster and easier. In 2021, the drone’s mainframe design embraced a two-layered structure with additional mounting holes and frame capacity. The triangular-shaped legs have multiple contact points to the front and back arms. The extra frame space improves customizability without redesigning the entire frame. Meanwhile, the curved legs make the drone rigid and resistant to impacts and pressure.

The software system incorporates conventional deep-learning algorithms for recognizing edges, shapes, etc. Many simplifications were made using OpenCV as one of our main software frameworks. For example, 3D data processing was done through an off-the-shelf depth camera from Intel. The built-in software solutions fully convert software commands through GPS and orientation data, which is the foundation of Nostromo’s behavior.

To solve IARC Mission 8, Ascend NTNU summoned the spirits of four legendary Pac-Man ghosts from 2018 to 2019. All Pac-Drone featured carbon fiber ducts around the propellers, providing a 20% increase in lift capacity and ensuring human-safe operation. The rest of the drone frames used carbon fiber and 3D printed parts to ensure strength and lightweight.

Each drone is equipped with an Nvidia Jetson TX2 computer. These provide ample computing power to handle data from stereo cameras and other sensors. The systems were built on the framework of ROS (Robotic Operating Systems), allowing for a modular structure of the software stack from the sensors to the control systems.

All four drones combined custom SLAM (Simultaneous localization and mapping) systems with data from stereo cameras to localize QR-code segments and human poses for gesture recognition. Field operators can command drones to perform complex tasks using simple speech recognition. Every command operation enables the built-in obstacle avoidance systems to avoid physical and environmental collisions on the field.

This year's drone is a new and improved version of last year´s drone. It got more computing power and sensors than the previous one. The drone's structure is made out of carbon fiber and 3D-printed parts. This makes it modular, which allows easy testing of multiple technologies.

The drone feature two Nvidia TX2s, which run the operating system based on ROS. The landing gear has multiple sensors for detecting height, landing, and physical contact with ground robots. The new and lighter cameras are isolated from the frame to minimize vibrations.

Our third quadcopter was designed as a physically robust platform for testing new control software. The small size and low weight mean that we can try custom-made frames made of carbon fiber and 3D-printed parts, allowing compact hardware placement and even weight distribution. In addition, well-balanced motors from T-Motor minimize vibrations and yield new control strategies, including landing on ground robots, with less risk of damage to equipment.

It has a Pixhawk flight controller and an Intel NUC onboard computer running Ubuntu Server with ROS. The custom-made frame made of carbon fiber and 3D-printed parts allows compact placement of the hardware and even weight distribution, and well-balanced motors from T-Motor minimize vibrations and yield high efficiency.

Our second aerial robot is custom-designed using carbon fiber and 3D-printed parts. It was built for the 2016 IARC and designed with Mission 7 in mind, and it can carry all the equipment we need to herd the target robots across the green line.

The sensors it has for navigation, ground robot detection, and collision avoidance are five cameras, one laser rangefinder, one 2D laser scanner, and an inertial measurement unit.

The data is processed using an onboard Intel NUC and an external computer communicating using WiFi. The flight controller is a Pixhawk connected to the
NUC.
‍
Our software can recognize the lines and corners of the grid, detect target robots and decide when and how to interact with them.