- Chiranjib Bhattacharyya (RBCCPS/Computer Science and Automation)
- Bharadwaj Amrutur (RBCCPS/Electrical Communication Engineering)
- Raghu Krishnapuram (RBCCPS)
- S. N. Omkar (Aerospace Engineering)
Civilian use-cases of drones is poised for exponential increase in the coming years, which is expected to generate significant economic value in several sectors such as agriculture, infrastructure assessment, law and order, journalism, etc. Central to all these sectors is the deployment of drones that can navigate autonomously in both structured and unstructured environments.
The goal of our research into autonomous navigation of drones is to understand the complexities of navigation of small quadcopters, with the aim of developing readily-deployable technologies for several application verticals. One of the scenarios we are exploring is the use of smart infrastructure to aid navigation. The premise is that if drones are allowed to interact with smart infrastructure, then we can achieve precision navigation without GPS by employing machine learning. The key challenge here is to put up minimalistic infrastructure, and yet be able to accomplish navigational tasks that require precision. This approach is highly interdisciplinary, and it brings to bear expertise in machine learning, sensor technologies, as well as drone design and construction to solve complex navigational problems. Some of our projects in this area are described below.
Autonomous outdoor navigation
The goal of the project is to achieve precise autonomous outdoor navigation of a drone without using GPS (i.e. the “under the tree canopy” scenario), between two waypoints inside the IISc campus, at an elevation of 2-3m from the ground. The approach uses the network of roads inside the campus and leverages our autonomous road following technology that uses the drone’s monocular camera. Road junctions are disambiguated with Smart City infrastructure assistance. Similarly, precision landing on predefined pads at the destination can be identified and characterised either by either smart-city infrastructure assistance or visual markers. The overall path planning for autonomous navigation within the IISc campus is achieved by using an Android application that has a database of Smart City infrastructure-assisted geofenced junctions, static obstacles, as well as landing and take-off pads.
The goal of this project is to enable an off-the-shelf drone to sense and avoid obstacles in unstructured environments using only a monocular camera. A depth map is essential for obstacle avoidance and path planning. The challenge lies in the fact that monocular vision is inherently deficient for depth estimation. To overcome this limitation, we have trained a deep neural network architecture for depth prediction from RGB images, with data collected inside the IISc campus using an RGB and depth sensor. The predicted depth maps are used for trajectory generation, followed by a controller which commands the drone to follow the trajectory. Currently, a clustering algorithm is being employed to generate single-step control actions such as “turn-left”, “go-straight” and “turn-right”. Other methods like deep reinforcement learning, and conventional path-planning techniques in 3D space are being explored.
Ultra-wide band transceivers for smart infrastructure
One of the technologies that is being explored for creating smart infrastructure for drone navigation is ultra-wide band (UWB) transceivers. We have demonstrated that high precision (less than 10cm error in lateral position) for localization and fast refresh rate (40Hz) can be achieved with UWB. Infrastructure-aided navigation is crucial in regions where uncertainties associated with fully autonomous flying may pose a risk. When compared with QR/AR Tags having waypoints embedded into them, UWBs offer more reliable operation even in low visibility conditions and can function as far away as 50m from the infrastructure. Other scenarios being explored in conjunction with UWB include facilitating bundle adjustment of SLAM-based. An added advantage is that any other information obtained through the infrastructure (such as cameras mounted on street lights) can also channelled to the drone through the cloud, aiding in global or local planning strategies.