Distributed Reconnaissance And Mapping system (DREAM)

The goal pursued in this study is to develop a robust system made of multiple unmanned ground robots to discover and map unknown places and environments that are potentially dangerous for humans, while being controlled from a VR environment. An example of a use case scenario would be the damaged airport of Zaventem during the attacks of 2016, where structural integrity of the building or the absence of additional explosives were not guaranteed. Navigating in such places might require interacting with objects (e.g. moving debris or opening doors). Therefore, we propose to equip our platforms with a robotic arm to allow these basic manipulations. The multi-agent setup allows redundancy (thus robustness to failure or destruction) and an increased mapping speed and manoeuvrability.

RMA will work on 4 different aspects within DREAM:
Multi-agent 3D mapping. the multi-agent configuration of SLAM poses several challenges and open questions, such as determining whether to let a central node handle the entire computational load or distributing equally among all agents. With this choice will come many subsequent challenges (telecommunication, adaptations in the SLAM pipeline, etc.). We propose to develop a unified approach for each building block of the real-time 3D SLAM in the multi-agent setup, which will be trained in a synthetic environment.
Generic robotic arm manipulations. The main challenge with robotic arm manipulations is to make sure that they remain correct even though the problem configuration (door handle, object to grasp, robot position) changes from the training and evaluation configurations. To cope with this, we will use the framework of reinforcement learning (RL) in a synthetic set-up to make our arm manipulations more robust to different setups.
Ground robot autonomous navigation. Multi-agent path planning in uncontrolled environments with high-level directions from a human operator remains largely unexplored in the literature. Starting from simple components, we will gradually increase the “unstructuredness” of the environment and the autonomy of our system to propose an online path planning algorithm that can take in high-level human directions (using the VR hand controllers).
Shared world representation. Since XR is still a relatively young technology, such applications have not been studied extensively. Here, we will have to propose a variety of ways to interact from the VR console with the robots. Keeping the end users in the loop, we will make sure to build a distributed world representation that is both intuitive and powerful enough to allow for all the interactions (moving in the 3D map, controlling robots’ navigation, commanding manipulations).