Machine learning boot camp for next-generation field robots
Alongside industrial robotics, which continues advancing at pace, interest in ‘field robots’ is growing for professions who rely on repetitive, dangerous or distant manual operations. For example, mining robots can map flooded passages and analyse mineral concentrations, while agricultural robots can pick fruit and milk cows. However, for wider adoption, the manufacture and operating costs of field robots will have to be reduced, but their capabilities extended – principally in the areas of sensing, perception and control. “Industrial robots function in more structured environments, such as factories, so are easier to design and operate. Field robots have to adapt to more unpredictability to fulfil tasks,” say Hugo Silva, coordinator of the EU-funded DEEPFIELD project. “While data-driven deep learning strategies could help accomplish this, there are still limitations in how these techniques are applied in real-world robotic activities.” DEEPFIELD enabled early-stage researchers (ESRs) from the INESC TEC Centre for Robotics and Autonomous Systems, the project host, to explore a range of deep learning techniques to give field robots the autonomy and adaptability needed to operate successfully in challenging environments. Amongst noteworthy successes, the researchers demonstrated how deep learning methods could help detect marine litter from space using hyperspectral imaging systems, and used imaging technologies to enable robots to establish their position and orientation while underwater.
Finding the best deep learning algorithm for the task at hand
In support of the EU’s ambitions for ‘Europe’s digital future’, DEEPFIELD was designed to give its ESRs varied hands-on exposure to robotic deep learning techniques. Focusing on a particular robotic problem where deep learning could help, the researchers first collected direct data or used a simulation environment, then after data analysis experimented with various algorithmic solutions, before evaluating the outcome. “The range of deep learning approaches and tasks to be accomplished was broad enough to give us a good grounding in a variety of solutions to build adaptable next-generation robots,” explains Silva. Throughout, the researchers worked with many kinds of robots: ground-based autonomous cars, unmanned aerial vehicles (UAVs), underwater remotely operated vehicles (ROVs) and autonomous underwater vehicles (AUVs). The researchers visited the facilities of consortium members: the University of Girona (Spain), Heriot Watt University (Scotland), the Max Planck Institute (Germany) and Politecnico di Milano (Italy), where each organised thematic workshops, summer schools, short scientific missions and mentoring. “After spending time learning the basics and working with experts in the specific field at a partner’s facilities, we tested software or algorithms using real robots in the summer schools,” says Ana Paula Lima, project manager. For example, the University of Girona focused on underwater manipulation, with the researchers learning about coding relevant to the Stonefish simulator, which they then used to perform inspection manoeuvres using an AUV/ROV.
In pole position for next-generation field robots
The project also gave the ESRs ample opportunities to network for their own professional development and to extend INESC TEC’s knowledge sharing with key players in the field. This gave the team a chance to present their work. For example, on a visit to the Arctic Pole’s NY-Alesund Research Station, Svalbard, two INESC TEC robots were tested, demonstrating the team’s capacities. “The feedback was so encouraging that a few weeks later INESC TEC submitted a proposal to the EC, which included new contacts made during that visit,” notes Lima. Having a more qualified and motivated young team has proven a boon to INESC TEC in its pursuit of additional projects in areas ranging from defence and security to the environment and offshore energy resources, where the researchers are now ready to apply their newly acquired knowledge.
Keywords
DEEPFIELD, algorithm, machine learning, robots, deep learning, autonomy, data, underwater, litter
