Research
Automatic Waste Management Plants
Ever wonder how robots are able to find and localize objects in recycling facilities? Object segmentation is the answer! We have collected a new dataset for segmenting objects using a combination of regular cameras and hyperspectral imaging, special sensors that can see more than meets the eye. This data combination improves the identification and localization of challenging objects present in waste more accurately, which means cleaner recycling and less amount of material ending up in landfills.
Give me one second and one video... From just observed demonstrations, we seek to learn the distributed scalable control policies that mimic a task. Thus, a team of robots can be deployed in time-varying scenarios, where the number of teammates or neighbors changes. We also propose a solution to learn, from those demonstrations, the underlying graph describing the interactions between robots.
"What is happening on the street? How many people are in the store? My child is lost in the shopping center, can you help me?" These are common questions in real-world scenarios that require scene understanding for monitoring or security. As a result, multi-camera systems are becoming increasingly prevalent in public spaces. However, processing the large amount of data they generate is a titanic task. To address this problem, we propose different approaches to automate the high-level understanding of scenes involving people by leveraging perception systems.
Neural networks have come a long way in extracting high-level knowledge from images of the world. However, there is a problem: they are usually overconfident about their predictions and are not good at expressing uncertainty, resulting in wrong information. Fusing several measurements is a good solution to improve the reliability of knowledge. In order to do that, we propose a method that correctly characterizes the noise and uncertainties of our "sensor", the neural network.
Are robots conquering Hollywood? It looks like that! Drones are more and more used in the cinematography industry and the solutions to make those drones autonomous are growing exponentially. We have developed different approaches to make drones more autonomous in the field of cinematography as well as some tools to help in the process.
Using Foundation Models to Perform a Drone Show
With the revolutionary emergence of foundation models, the possibilities for their application continue to expand. In this solution, we leverage a foundation model as feedback to control a swarm of drones, creating drone shows that display various shapes with just a single word as instructions. Want to learn more about how this innovative technology works? Click 'Read more' and unlock the secrets behind our mesmerizing drone displays.
Stop!!! I can't catch them!! The herding problem consists in driving a group of targets to specific locations by coordinating a team of robots. Particularly of interest is the design of control strategies which outcome complex nonlinear systems, with targets described by noncooperative reactive behaviours. This problem archetype is extensible to a wide variety of real-life problems such as monitoring of in-danger species, fire extinction or evacuation protocols.
We need to be fast, the battery is draining! In order to improve the efficiency of scene coverage with drones, we first capture a coarse map of the scene in real-time. Then, we use this map to identify regions that were not correctly observed and generate sweep trajectories to observe them at once. The trajectories are distributed to a team of drones that executes them in a coordinated way, minimizing the total mission time.
The people united will never be defeated. In this research line we find optimal ways of recovering the global estimates of quantities of interest by means of distributed algorithms that do not need any global knowledge
It's fast, it's guaranteed and it's distributed! We present GeoD, a consensus-based distributed SE(3) pose graph optimization algorithm with provable convergence guarantees. This algorithm enables a group of robots, given noisy relative pose measurements, to reach agreement on the group’s SE(3) pose history in a distributed manner.
Tag your target and don't let it out of your sight!! Cooperative observation of multiple moving objectives requires for each robot to constantly re-decide to whom it should observe. This should be done ensuring that all objectives are perceived at all times, but local communications and the highly dynamic nature of everybody's motion makes this task assignment problem particularly difficult. In the team we have been able to devise robust and computationally light distributed algorithms that efficiently solve dynamic task assignment problems.
Environmental monitoring, security and surveillance or dust cleaning are repetitive tasks perfectly tailored for the use of multi-robot systems. Persistent coverage offers a general framework to address these problems in a compact and formal way. We have developed several perception, control and planning solutions where teams of robots efficiently execute a persistent coverage task in complex environments.
We have extensive experience desigining distributed algorithms for safe and coordinated motion of teams of robots in complex environments. Our methods consider onboard perception and account for static and dynamic obstacles in the environment.