Robots can often face challenges with communications and global positioning system (GPS) specifically in underwater and indoor environments. In this thesis, we address the hardware and software challenges for mobile robots operating in communications constrained and GPS denied environments.
An example underwater environment scenario is coral reef and marine life monitoring missions which may last hours and generate a long footage with limited to no bandwidth communications to the surface. Thus, we propose a solution to sample the data efficiently. To do so, a capable hardware needs to be installed first. Using the MARVL-ROV underwater robot, which we developed, we are able to run MERLION, a modern, semantically aligned learning-based visual (camera) data sampling in real time onboard the robot during the mission, adaptable to the user’s prompt of topic of interest. This saves only the most relevant, representative and unique frames to ease storage space or for real-time low-bandwidth live streaming to the surface vehicle. A particular example of low-bandwidth underwater communications is cutting edge acoustic and optical untethered communications between the underwater robot and surface vehicle. MARVL-ROV with our proposed MERLION framework has been deployed at different islands in Singapore with success rate of 70% on average compared to human level of semantic efficiency.
For indoor environments, some example scenarios without communications and GPS that we consider are search and rescue and emergency evacuation. In such cases, we proposed two useful solutions. These include local scene graph generation and robust autonomy through simulation. The real-time generation of 3D spatio-semantic scene graphs in field robots has been of great interest in recent years, allowing them to have a high-level and hierarchical representation of the scene at hand for use by all modules such as reasoning, planning, perception, and navigation. We develop a bimodal fast real-time 3D scene graph generation algorithm for RGB-D input for open-set tasks with competitive performance compared to the state of the art. We implemented in real world on Jetson AGX Orin, allowing robots in indoor environments to have greater latitude and control in decision planning especially autonomous scenarios, thereby requiring little to no communications with remote servers or other robots. Our Habitat simulation benchmark testing showed faster task convergence timing on some open-set tasks compared to the state of the art, and three times faster in scene graph generation time for the fast mode. In addition, we also developed the Astralis simulator for robust obstacle generation for robust testing of robotic algorithms in simulation to allow autonomy algorithms to be developed with higher reliability in messy indoor environments
- Intelligent Mobile Robots in Low-Bandwidth Indoor and Underwater Environments: A Journey from Designing to Real-world Deployment
- Marcel Bartholomeus Prasetyo
- Malika Mehdi Mohmed Meghjani - Singapore University of Technology and Design, Information Systems Technology and Design (ISTD)
- Singapore University of Technology and Design; Master of Engineering (Research)
- Master of Engineering (Research) , Singapore University of Technology and Design
- 86
- 9914130609846
- Information Systems Technology and Design (ISTD)
- English
- Thesis