Master thesis proposals - external

Ericsson Research

Looking for a thesis project during Spring, starting in January 20223? Did you know Ericsson develops research on themes related to robotics and AI? Take a look at the current projects we have open:

Project 1: Assisted teleoperation

5G and 6G technologies are pushing boundaries of robot control over network connections. The purpose of this project is to apply concepts of state-of-the-art model-based control methods to robots being operated remotely. The goal is to assist the user and provide safety guarantees in the presence of dynamic obstacles, communication delay and possible packet loss. The project can be performed in simulation and/or on real hardware. Applicants should have a solid background in control theory (such as MPC and control barrier functions) and robotics.

Project 2: Collaborative SLAM

In this project you will investigate optimal strategies for collaborative simultaneous localization and mapping (SLAM) using modern smoothing and mapping techniques with graphs as the underlying structure. In particular, the goal is to investigate how collaborative SLAM can be performed in such a way that resource-constrained robots can achieve high performance, while minimizing the computational burden. Applicants should have a solid background in SLAM and robotics.

Project 3: Topology-Aware Robotic Exploration

Autonomous exploration algorithms allow robots to discover and map an environment. However, state-of-the-art approaches are information greedy and often give disorganized motion patterns. To increase efficiency, this project will look into exploiting knowledge about the topology of the environment, which can be used to give priorities to certain regions. Applicants should have a solid background in robotics and algorithms, including simulation tools like ROS and Gazebo.

Project 4: Edge-Assisted Object Detection

In this project you will work with an existing testbed for distributed object detection, where the computations can be offloaded from devices to edge servers. This requires a scheduler, that decides what computations to run and where. In one sub-project we will study a deep learning-based method to improve the performance of the current scheduler, along with exploring the impact of some network metrics on the offloading strategy. In another sub-project, we will study the effect of video content on the object detection performance and the offloading strategy. Applicants should have a solid background in ROS, computer vision and/or deep learning.

Requirements

All the projects require some common skills:

• Strong ability to formulate problems and solve them, independently and in groups.
• Programming skills, preferably in Python or C/C++.
• Strong communication skills in written and spoken English.

To apply or get more information

Project 1: Roberto Castro Sundin - roberto.castro.sundin@ericsson.com

Project 2: David Umsonst - david.umsonst@ericsson.com

Project 3: Fernando dos Santos Barbosa - fernando.dos.santos.barbosa@ericsson.com

Project 4: Alejandra Hernández Silva - alejandra.hernandez.silva@ericsson.com

Ocean Infinitiy

Ocean Infinity has a number of proposals  for thesis projects during Fall, starting in the end of August 2022.  These involve problems within underwater perception, sonar and robotics.  For example:

• Target Identification and classification from sidescan sonar  data.
• Marine Growth assessment using computer vision, (Marine Growth Classification)
• MultiBeam Echo Sounder (MBES) image alignment (Multibeam Sonar Image Alignment)
• Image quality assessment (visibility, exposure, lighting, positioning etc) (Automated Image Assessment)
• Localization of AUV swarm (SLAM loop closing)

Please contact John Folkessn johnf@kth.se for more information.

  

Arriver Sweden

Knowledge Distillation for Autonomous Driving

Description

CNN-based association for Multiple Object Tracking

Description

Scania

Localization For Autonomous Driving in Highway Scenarios using Statistical Filtering

More details

Landmark Detection for Autonomous Vehicles using Machine Learning Techniques

More details

Efficient Map Generation and Update for Autonomous Driving

More details

GPU-accelerated Multi-Camera Visual Odometry for Autonomous Driving

More details

Data-driven motion planning for heavy duty vehicles

Autonomous driving technologies are expected to bring fundamental changes in near future, including improved traffic safety, more efficient and accessible mobility. Motion planning is the process of generating feasible, collision-free and optimal paths in real-time. While the definition of feasible and collision-free are clear and given by physical equations, optimality depends on a multitude of factors including, but not limited to, vehicle type, application domain, user preference.

Motion planners use a cost/reward function to define optimality, which then dictates the behaviour of the autonomous vehicle. While it is possible for an expert to manually design and tune such a function, it is a tedious process. The goal of this master thesis is to investigate how to leverage data obtained from human drivers to tune the reward/cost function for heavy duty vehicles through an automated process.

Application link, full description and details: https://www.scania.com/group/en/home/career/available-positions/job-description.html?q=*&job_id=19291&lang=en

Univrses

Cross-city domain shift is the cause of a 25-30% metrics drop of normal Deep Learning predictors. Change of weather and lighting conditions might even result in more disruptive effects. Therefore, to provide a robust and accurate output, the predictor would need to be trained on every possible deployment condition, e.g. multiple cites, districts, seasons, kinds of weather, cameras pose and intrinsics. This largely reduces the scalability of AI solutions both in terms of costs and deployment time. Recent Unsupervised Domain Adaptation (UDA) methods proved to be very effective in mitigating (or even solving) these shortcomings.

Contact: Pier Luigi Dovesi <pier.luigi@univrses.com>
Application link, full description and details: https://career.univrses.com/jobs/1401972-master-thesis-project-unsupervised-domain-adaptation

ABB Corporate Research

Using dialogue to disambiguate robot manipulation, learning from demonstration.

Behavior Trees are a reactive task switching policy representation, used to control robotic agents. LfD can be used to teach the robot a task and BTs can be generated out of human demonstrations. However, ambiguities might rise if the target object for the task is similar to other objects in the environment. Thus, verbal-HRI can be used to disambiguate the task, both during the learning step and the execution step. The objective of the project is then to build a framework for continuous interaction between human and robot, from learning a task to execute it.

Start: between Jan. 2022 and June 2022
Duration: 6 months
Place: ABB CRC (Västerås)
ABB will cover the accommodation in Västerås
Contact: Matteo Iovino, matteo.Iovino@se.abb.com

Details here.

SkyCraft AB

Master projects in CV and ML on powerline detection

Description

Viscando AB

Projects within deep learning, signal processing and modelling for traffic and autonomous vehicle safety

Description

Babyshop Group

Leveraging transfer learning for multi-label attribute prediction in fashion images, using deep learning

At Babyshop Group, there exist tens of thousands of images of children’s apparel. Currently, new items are labeled by hand, where each item has multiple attributes, eg. color, category, pattern, neckline, and style, etc. Since the workload of this manual task is immense, a need for a support system is high. In literature, a multitude of previous work has been done in regards to multi-label attribute prediction on fashion clothing, most notably DeepFashion and iMaterialist Fashion. Here, deep learning models have been trained on hundreds of thousands of labeled images to predict corresponding attributes, which has proven highly successful. However, almost all of these images are of adult fashion, with very little emphasis on childrens’ fashion. The question then arises, is it possible to combine the pre-trained models from previously mentioned multi-label attribution prediction models that are trained on almost exclusively adults’ fashion, together with the learnings from transfer-learning frameworks - in order to successfully perform multi-label attribution prediction on childrens’ fashion?

Required qualifications: MSc studies in Computer Science, Machine Learning, Computer Engineering, Mathematics, Physics, or related field; Good understanding of machine learning frameworks such as Keras, TensorFlow, PyTorch, Scikit-Learn, and/or Spark; Proficiency in Python and Git; Knowledge about data wrangling and data munging, using SQL, Pandas, and Numpy.

Web version of the proposal with instructions on how to apply.

Time frame: from early January to mid-June.

Contact: Marcus Svensson (marcus.svensson@babyshop.se), Data Scientist at Babyshop Group

Scania Autonomous Transport Solutions

A thesis project at Scania is an excellent way of making contacts for your future working life. Many of our current employees started their career with a thesis project. 
The concept of autonomous driving has rapidly transitioned from being a futuristic vision of robotics and artificial intelligence to being a present-day reality. Scania is among the companies aiming to sell fully autonomous vehicles a few years from now, but many research challenges still needs to be solved.

Within the topics of robotics, perception and learning we have several interesting projects:

• Risk-aware Decision Making with POMDPs
• Online Estimation for of Safety Margins for Risk Assessments 
• Behavior Prediction of Surrounding Cyclists for Autonomous Driving
• Collision checking for articulated buses
• Steering Methods for Nonholonomic Motion Planning

More projects may also be published at Scania Career.

Contact:
See link for each project, or contact Truls Nyberg, Industrial PhD Student Scania/RPL
trulsny@kth.se