Internship - Robotics and Automation Unit

Eurecat is currently the leading Technology Centre in Catalonia, and the second largest private research organization in Southern Europe. Eurecat manages a turnover of 50M€ and 650 professionals, is involved in more than 200 R&D projects and has a customer portfolio of over 1.600 business clients. Eurecat is currently participating in more than 60 EU funded collaborative projects, mainly in the Horizon 2020 Programme. In addition to this wide experience at European level, Eurecat is also a strong player in the various R&D programmes sponsored by the Spanish and Catalan administrations. Technology transfer is also an essential activity in Eurecat, with 36 international patents and 7 technology-based companies (eight in Spain and one in Latin America) started-up from the centre.

is devoted to performing research and new developments in the field of robotics. The current research areas are: 1) Robotics Manipulation, 2) Cognitive Robotics, and 3) Mobile Robotics. In the last 15 years, the group has executed a total of around 125 research projects including applied research (15 European, 20 National) as well as technology transfer with over 90 private projects focusing on Advanced Robotics Applications with industry. The group is active in different domains including Agile manufacturing, Smart logistics, Infrastructure Inspection & Monitoring, Environmental, Agriculture and Construction.

This year, the Robotics Unit is offering different proposals to be developed as Master Thesis and/or Internships, which are listed below. If the student has its own proposal, it also could be taken into consideration. All proposals will be carried out at Eurecat’s HQs in Cerdanyola del Vallès https://maps.app.goo.gl/bb32tGvdrhua1Z8X6  where the Robotics department is located.

Ms. thesis proposals 

• Robotics Manipulation
  • Development of safe learning strategies based on constraint manifolds
    • Reinforcement learning in robotic agents is particularly complex due to several challenges including safety during the learning process and mechanical constraints. These challenges are not commonly addressed in the generic artificial intelligence literature since they do not deal with physical agents. In this project, common reinforcement learning strategies will be combined with constrained optimizations in the constraint space to ensure secure learning.
  • Diffusion policies: Applying Diffusion Models to Policy in Robot Learning
    • This thesis aims to explore the application of diffusion models to policy learning in robotics. Students will investigate how diffusion models, a class of probabilistic models, can be used to represent and learn complex policies for robots. The research will focus on developing efficient and scalable methods for policy optimization, reinforcement learning, and decision-making in various robotic tasks. The thesis will contribute to advancing the understanding of diffusion models in the context of robotics and their potential for improving policy learning and decision-making processes.
  • Learning policy and reward function using active reward learning techniques minimizing human supervision
    • In robot learning contexts, defining a function that specifies when to reinforce the agent's actions can often be very expensive. To address this challenge, algorithms will be developed that will not only learn the robot's control policy but also learn the reinforcement function. To do this, supervisors (human or based on conventional task planners) will be used to evaluate the robot performance. Strategies will be implemented that will minimize the number of supervision calls and maximize the information obtained in each call.
  • Implementation of randomization strategies for learning conditions to minimize the Sim2Real gap
    • This project addresses the challenge that exists to be able to take advantage of learning conducted in a simulated environment in real environments. Mechanisms will be introduced for a greater generalization of the control policies learned, focusing on the randomization of physical conditions of the simulation environment. These strategies will be applied transversally to the different learning methodologies so that it is expected to obtain robust learning to variations in the conditions of the real environment.
  • Curation Strategies for UMI Data (https://umi-gripper.github.io/) for Target-Domain VLA Adaptation
    • Objective: Create a dataset curation strategy for public UMI data to finetune VLAs.
    • Key tasks:
      • Gather a dataset with the UMI gripper for a target domain.
      • Develop strategies to add language instructions to UMI data.
      • Develop curation strategies to identify and combine optimal subsets of public UMI datasets, enabling a pre-fine-tuning stage that yields robust target-domain fine-tuning.
      • Develop and apply validation methods (coverage, density, and quality checks) to assess curation effectiveness.
    • Finetune a VLA for the target domain and deploy it on a real robot
  • Egocentric Human Data Collection for Robot Learning with Gaze-Aware Attention Models

    • This thesis focuses on developing a pipeline for capturing egocentric human demonstrations for robot learning using a wearable setup with head-mounted and eye-facing cameras. The objective is to record both hand-object interactions and human gaze during task execution, enabling the study of visual attention as an additional supervision signal for robot learning. The student will investigate methods to retarget these demonstrations into robot-compatible data representations and evaluate their use for training manipulation policies in egocentric imitation learning. The final goal is to assess whether gaze-aware demonstrations improve policy learning, generalization, and task-relevant attention in robot manipulation.

  • Dexterous In-Hand Manipulation
    • This project focuses on developing an AI-powered system for an anthropomorphic robot hand equipped with tactile sensors. While initial work has successfully enabled the hand to grasp objects firmly, maintain a secure hold, compensate for external forces, and rotate objects (e.g., cubes); the next steps will investigate and implement novel Reinforcement Learning (RL) algorithms using the existing IsaacLab Sim2Real pipeline. The primary goal is to extend the existing pipeline to include novae AI/RL models, vision and/or tactile based approaches pushing the boundaries of dexterous manipulation beyond stable holding to dynamic reorientation. Students will explore advanced RL techniques to achieve robust and adaptive control for complex in-hand manipulation tasks.
  • Machine Learning for Pose Estimation and Control of Deformable Linear Objects
    • This thesis will focus on developing machine learning-based methods for the perception and manipulation of Deformable Linear Objects (DLOs) such as wires, cables, ropes, or shoelaces. The project will explore pose estimation and shape reconstruction techniques using sensory data (e.g., vision and tactile feedback), leveraging Principal Component Analysis (PCA) and deep learning architectures to efficiently model the complex, non-rigid deformations of DLOs. The main objective is to enable robots to recognize, predict, and manipulate DLO configurations with precision and adaptability in real-world scenarios.
  • Multi-robot Manipulation

    • This thesis will investigate the coordination and collaboration of multiple robotic agents to achieve complex tasks that are difficult or

      impossible for a single robot. Students will develop algorithms and strategies for task allocation, motion planning, and inter-robot communication to enable efficient and robust collaborative manipulation, assembly, or other multi-robot operations. The research will contribute to advancing the capabilities of multi-robot systems in various applications, from industrial automation to complex environmental exploration.

  • Screwing and Unscrewing application for battery pack assembly and electrical recycling
    • This thesis focuses on the development of unscrewing and screwing application. More precisely it requires the development of a reliable and flexible pipeline for un/screwing application in the field of battery pack recycling based on visual servoing. The student is required to evaluate the current approach to solve the task, evaluate possible alternative pipelines to increase the success rate, and validate the solution in a real car battery pack.
  • Intelligent assembly of non-standard interlocking parts
    • This thesis focuses on the automated assembly and kitting of components with irregular, interlocking geometries. While traditional robotics handles standard shapes well, the assembly of non-standard parts (e.g., snap-fit panels or tool shadow boards) still presents a challenge in terms of geometric reasoning and precision insertion. The project requires the development of a vision-based pipeline to identify irregular parts, estimate their 6-DOF pose, and perform precise placement into matching slots using the adequate tool. The student is required to evaluate current shape-matching techniques, develop a "search-and-fit" insertion strategy to handle tight tolerances, and validate the solution in an industrial scenario
  • Optimal Next Best View Planning

    • This master thesis focuses on developing and implementing a Next Best View (NBV) planning algorithm designed to optimize the sequential acquisition of observational data for a scene reconstruction task. The primary objective is to strategically determine the optimal next sensor pose (position and orientation) that will yield the "best" expected improvement in the current understanding or model of the observed scene.

  • Collaborative robot teaching using AR/VR
    • The objective of the thesis is to create and implement a teleoperation system assisted by augmented/virtual reality devices to command a collaborative robot with the purpose of improving the robot performance and teach it new skills.
  • Photorealistic simulation using Unreal engine, Gazebo and ROS
    • While many efforts are put in the precision of simulations in terms of motion, visual representations are often pushed to the background. This limits the reliability study of camera-based algorithms, especially those depending on visually plausible data. AIM: study state-of-the-art simulators with photorealistic visual capabilities for the assessment of mobile robotic navigation algorithms which depend on accurate and realistic environment information.
      • A literature review of state-of-the-art solutions
      • The definition and implementation of sensors, and sensor and system dynamics
      • The creation of a specific outdoor environment in line with applications currently ongoing at Eurecat
      • The implementation of simple navigation algorithms to validate the proper working of the simulation
  • Adaptive Human-Robot Interaction Using Personalized Vision-Language Models
    • This thesis explores how personalized vision-language models (VLMs) can enable robots to adapt their responses and behaviors to individual users during social interaction. The project involves integrating a user-aware VLM with a real robot, allowing the system to perceive visual and linguistic cues and adapt in real time. The research focus can be on either (a) making robot interactions more expressive and socially attuned to the user, or (b) improving the robot’s ability to follow user-specific instructions. Key challenges addressed will include bias mitigation, ethical personalization, and real-time system performance. The aim is to assess how personalized VLMs can make robots safer, fairer, and more effective collaborators for diverse users in real-world scenarios
  • Machine learning for screw detection and classification
    • This thesis focuses on the problem of screw detection ad classification for recycling and automotive application. In this project, a full machine learning pipeline will be developed starting from the dataset collection and labelling, the selection of a machine learning model, the training, and its evaluation in an industrial relevant scenario. The final objective is to have an ML model capable of localizing and classifying the different screws present in the robotic demonstrator.
  • Mathematical modelling and uncertainties estimation of a robotic end-effector
    • This thesis focuses on the development of a mathematical model of a robotic end-effector. The tool is composed of a screwdriver used for fastening applications. The presence of a flexible joint to reduce the vibration, however, it reduces the precision of the position and makes fastening and unscrewing operation impossible without additional control. In this project, the student is required to mathematically model a mass-spring-damper system of increase complexity to reproduce and model the tool uncertainties to permit a precise localization of the tool tip.
  • Hyperspectral Imagin in Industrial Application
    • This thesis focuses on the application of hyperspectral imaging for industrial applications. Hyperspectral imaging (HSI) is a sensing technology that captures both spatial and detailed spectral information across hundreds of narrow bands, ranging from visible to infrared light (250 nm–15,000 nm). For these reasons, the representation offered by these cameras is richer and enables a more easily understanding of material properties. This enables the camera to be used for a large set of applications including defect detection, material identification, and object detection. In this thesis the student is required to study the feasibility to apply these technologies to real-world industrial application
  • Robotic Learning for Wiring harness Assembly
    • This thesis focuses on the development of learning-based robotic manipulation policies for complex electronic component assembly tasks. These tasks, such as connector plugging and unplugging, are characterized by tight tolerances, contact-rich interactions, and high sensitivity to misalignment, making them challenging for traditional model-based or purely vision-driven approaches. The thesis explores the integration of multimodal sensing, combining visual information, robot proprioception (joint positions and torques), and tactile feedback. The student is required to investigate the feasibility of applying these methods to real-world industrial scenarios. Experimental validation will be carried out on automotive connector assembly tasks, as well as in recycling contexts where variability in object geometry and condition introduces additional uncertainty.
  • Computer vision for robotic agricultural application
    • This thesis focuses on the application of panoptic segmentation techniques for agricultural robotics. Panoptic segmentation is a computer vision task that unifies instance segmentation (identifying individual objects) and semantic segmentation (assigning a class label to each pixel), providing a comprehensive scene understanding. This representation is particularly relevant in agricultural environments, where robots must distinguish between multiple objects such as fruits, branches, and leaves, often under cluttered and unstructured conditions. The thesis explores the use of deep learning models for panoptic segmentation, leveraging multimodal visual data and state-of-the-art convolutional or transformer-based architectures. A key aspect of the work is the creation of a dedicated dataset tailored to agricultural applications, including tasks such as pruning and fruit harvesting, which require precise identification and localization of targets. The student is required to collect and annotate a dataset in real or simulated agricultural scenarios, train and evaluate a machine learning model for panoptic segmentation, and assess its applicability in downstream robotic tasks. These include motion-free robot planning, pruning target selection, and fruit grasping, with the objective of evaluating how improved scene understanding can enhance decision-making and manipulation performance in agricultural robotics.
  • Microphones in Robotic Assembly Tasks
    • This thesis explores the use of new modality to acquire information related to a robot execution task. In particular, is related to the use of “sound” and how this can be used in robotics to understand the environment. Development of a data acquisition pipeline for the use of a USB microphone in a robotic end-effector. The microphone is used to acquire information about the successfulness (e.g., the “clip” of an ethernet connector) or the quality of the task execution (e.g., sliding on different surfaces). The student is required to include the sensor in a robotic gripper and develop the necessary pipeline for data acquisition and filtering of the environment noise. The development can include the training of machine learning model for binary classification.
  • Robotic target tracking in industrial conveyor belt
    • This thesis focuses on the application of real-time target tracking techniques for industrial conveyor belt systems. The thesis explores the use of deep learning models or Kalma-filter approaches for visual tracking using a camera as reference. The student is required to work on an industrial conveyor belt robotic application developing the platform. Develop a tracking pipeline in real or simulated industrial environments. Optimize and evaluate an algorithm for robust target following and assess its applicability in downstream robotic tasks. These include high-speed pick-and-place operations, dynamic path plannin.

• Cognitive Robotics
  • Building Dynamic Knowledge Graphs from Multimodal Perception Streams in ROS2
    • This thesis focuses on bridging the gap between low-level perception (visual, audio, and depth sensing) and structured world knowledge for robotic reasoning. The student will design and implement a framework that translates continuous ROS2 perception topics — such as people/faces, asr_result, and skeleton_detection — into dynamically updated Knowledge Graphs or 3D semantic scene representations.
    • The work involves:
      • Integrating perception modules from the EutPerceptionStack (audio + visual + multimodal) into a unified world model.
      • Studying state-of-the-art techniques for spatial-temporal knowledge construction and incremental 3D scene updating.
      • Implementing methods for object-person-event linking and context persistence across robot sessions.
      • Evaluating the system on real robot platforms
  • Learning Priors from Human Videos
    • Objective: Develop a framework that translates predefined roles into actionable desires and dynamically prioritizes tasks based on contextual needs.
    • Key tasks: Utilize human video datasets to
      • Decompose complex activities into meaningful components.
      • Structure knowledge for reusability.
      • Define open-world constraints for adaptive learning.
  • Open-World Reactivity
    • Objective: Develop policies that enable immediate, context-aware reactions in open-world environments, with applications in Human-Robot Interaction (HRI).
    • Key tasks:
      • Learn policies for immediate, context-aware reactions in open-world environments.
      • Investigate potential applications in Human-Robot Interaction (HRI), leveraging human video data for training.
  • From Diagrams to Plans
    • Objective: Leverage diffusion models to enhance rollout strategies in Monte Carlo Tree Search (MCTS) for improved decision-making in stochastic environments.
    • Key tasks:
      • Explore how diffusion models can enhance rollout strategies in MCTS.
      • Assess the potential for improved decision-making in stochastic environments.
  • Anticipating Human Intent and Cognitive State through Multimodal Perception for Proactive Robot Collaboration
    • This thesis aims to develop a multimodal system that enables robots to anticipate human actions, interpret implicit commands, and adapt behavior in real time during collaborative tasks. By integrating motion trajectories, gaze direction, hand-object interactions, and physiological or behavioral cues (e.g., hesitation or repetition), the robot will infer short-term intentions and estimate the human’s cognitive state. Predictive models such as RNNs or Bayesian inference will be used to enhance the robot’s ability to proactively assist, modulate its behavior based on user workload, and understand context-rich, nonverbal instructions. The goal is to improve fluency, trust, and efficiency in human-robot teamwork through deeper perceptual and cognitive alignment.

• Mobile Robotics
  • AI for anomaly detection in tunnel images from drone images
    • Supervised learning is the basis for object/defect detection. It works well when you know what you are looking for, however what happens if you want to look for unforeseen objects? We propose to study anomaly detection algorithms, in particular autoencoder structures with segmentation that enables the detection and highlight of out of distribution objects.
      • A literature review on Anomaly Detection and segmentation
      • Implementation and adaptation of existing algorithms
      • Testing on real data
  • Non-linearMPC control on resource restricted drones
    • Drone control is usually tackled by applying cascade controllers. The inner loop stabilizes the aircraft attitude (roll and pitch), the yaw rate and the aircraft vertical velocity. By sending commands to the motor ESCs. On the top of it, an external loop controller oversees transforming position/velocity references in attitude commands. Although MPC is a well stablished technique, when applied to non-linear systems it can be computationally consuming. This work would try to devise which are the possibilities of implementing linear and non-linear in resource limited computers on-board of small drones to control the outer dynamics of the vehicle.
  • Integrating Large Language Models for Robot Goal Understanding
    • Objective: Explore the use of large language models (LLMs) to translate high-level human commands into actionable navigation tasks for a ground robot.
    • Key tasks:
      • Build an interface for processing and interpreting natural language commands.
      • Integrate with the robot's control system to execute tasks.
      • Validate in a scenario such as fetching and delivering items in a controlled environment.
  • Autonomous Crop Monitoring with Multispectral Sensors
    • Objective: Develop a ground robot system for real-time crop health monitoring using multispectral or thermal cameras.
    • Key tasks:
      • Integrate multispectral cameras with a ground robot.
      • Develop algorithms for detecting plant stress, diseases, or nutrient deficiencies.
      • Validate system performance in simulated or small-scale agricultural fields.
  • Role-to-Desire Mapping for Prioritization
    • Objective: Develop a framework that translates predefined roles into actionable desires and dynamically prioritizes tasks based on contextual needs.
    • Key tasks:
      • Develop methodologies for transitioning from defined roles to actionable desires.
      • Establish techniques to assign priorities dynamically based on contextual needs.
  • Neural Radiance Fields for Real-Time Mobile Robot Localization
    • Objective: Develop an efficient Nerf / Gaussian splatting system that allows mobile robots to perform simultaneous localization and mapping (SLAM) in real-time using neural representations. The system is expected be embedded in a edge device such as Jetson Orin for real-time operation in real robots.
    • Key tasks:
      • Data selection, environment modeling
      • NeRF implementation and training
      • Real-Time localization system design
      • Evaluation, benchmarking, field testing
  • Targetless Calibration Using Visual-LiDAR Correspondences for Non-Overlapping FOVs
    • Objective: Develop a targetless calibration method that leverages visual and LiDAR data correspondence, possibly using robot localization inputs from a SLAM system, for calibrating sensors with non-overlapping FOV equipped in a mobile platform, performing optimization or learning-based estimation of intrinsic and extrinsic parameters without the need for external markers.
    • Key tasks:
      • Feature matching algorithm between 2D image points and 3D LiDAR points
      • Framework to solve intrinsic and extrinsic parameters of the sensors
      • Calibration bencharking with real-world datasets or simulations
  • Enhanced Traversability Perception with Computer Vision and Machine Learning
    • Objective: Develop a computer vision module based on machine learning that integrates with an existing LiDAR-based traversability analysis system, enabling more precise classification of terrain elements such as non-traversable vegetation or traversable puddles, thereby improving the robot's autonomous navigation capabilities.
    • Key tasks:
      • Dataset Selection and Preparation
      • Learning Architecture Design
      • Integration with Existing LiDAR Pipeline
      • Model Training and Validation
      • Qualitative and Quantitative Evaluation
  • Frameworks and Methodologies for Realistic Outdoor Simulation in Autonomous Navigation
    • Objective: To conduct a comprehensive research study and comparative analysis of existing frameworks, platforms, robot models, and sensor simulation techniques for creating and utilizing realistic, unstructured outdoor environments to robustly evaluate autonomous navigation systems.
    • Key Tasks:
      • Comprehensive Literature Review and State-of-the-Art Analysis
      • Comparative Analysis and Benchmarking
      • Proposed Integrated Methodology and Best Practices
  • Passive Sensor-Based Localization and Local Navigation
    • Objective: To investigate the challenges and state-of-the-art in autonomous robot localization and local navigation using only passive sensors (e.g., monocular or stereo cameras, IMUs) in outdoor environments. This research will culminate in the design and initial implementation of a foundational system demonstrating these capabilities without relying on active sensing.
    • Key Tasks:
      • Design of an Initial Passive Navigation System Architecture
      • Proof-of-Concept Implementation of Key Modules
      • Preliminary Evaluation and Analysis
  • Cooperative Perception and Shared Maps in Multi-Robot Outdoor Missions
    • Objective: Explore techniques for sharing local perception data and maps across a fleet of robots operating in unstructured outdoor areas.
    • Key Tasks:
      • Communication and data synchronization strategy.
      • Merging LiDAR/vision-based local maps.
      • Evaluation of collaborative mapping accuracy and robustness.
  • Robust Road Surface Segmentation for Off-Road Navigation
    • Objective: Develop a lightweight semantic segmentation pipeline to distinguish traversable surfaces from vegetation and debris in off-road scenes using RGB-D or stereo input.
    • Key Tasks:
      • Dataset preparation or use of public datasets (e.g., RELLIS-3D, RobotCar).
      • Design and training of segmentation model (e.g., MobileNet, Fast-SCNN).
      • Integration into navigation stack and performance benchmarking.
  • Learning Terrain-Aware Locomotion Policies for Wheeled or Legged Robots
    • Objective: Train policies that adapt motion strategies to terrain types (e.g., mud, gravel, grass) to improve mobility in off-road or degraded environments.
    • Key Tasks:
      • Dataset creation with terrain labels and wheel/leg behavior.
      • Policy learning via reinforcement learning or supervised imitation.
      • Evaluation in simulation or with real quadruped or UGV.
  • Safe and Explainable Navigation Using Visual Question Answering (VQA)
    • Objective: Use VQA-based systems to explain robot decisions and query scene understanding during navigation missions.
    • Key Tasks:
      • Use of pretrained VLMs or LLM+vision models for scene Q&A.
      • Integration into robot interface for debugging or HRI.
      • Testing in synthetic datasets or controlled scenarios.
  • Energy-Aware Path Planning in Multi-Robot Systems with Shared Resource Constraints
    • Objective: Create a coordination framework that allows multiple robots to plan energy-optimal paths while sharing limited charging stations or re-supply depots.
    • Key Tasks:
      • Energy modeling and prediction based on robot type and terrain.
      • Prioritized or reservation-based access to shared resources.
      • Evaluation of system-wide efficiency and fairness.
  • Learning Traversability-Aware Navigation Policies from Raw Sensor Input
    • Objective: Train a navigation policy that adapts robot motion based on perceived terrain traversability from camera and/or LiDAR input, without relying on a global map.
    • Key Tasks:
      • Dataset preparation or use of traversability datasets (e.g., RELLIS-3D).
      • Learning-based policy (e.g., CNN+RL or transformer-based policy).
      • Real-time inference and reactive motion in unseen environments.