ETH Zurich : Computer Science : Pervasive Computing : Distributed Systems : Education : SSS FS2018

Smart Systems Seminar FS2018

Various topics from Ubiquitous Computing, Human Computer Interaction, Robotics and Digital Fabrication

Topic assignment

Time and Place

Goal

Organization

The seminar consists of talks given by students on selected topics and discussions led by the instructors. A maximum of 12 students will be admitted to the seminar. Priority will be given to Master students who have sufficient background knowledge in the topic but the seminar is generally open to Bachelor and Doctoral students as well.

Seminar attendees select a specific topic within the broader context of current research and prepare an oral presentation. As a starting point, the students are assigned 3-4 important papers in their topic but they have to collect complementary materials and compile them together. Oral presentations must be planned for 45 minutes. Each presentation will be followed by a technical discussion as well as a short feedback session on the quality/style of the presentation. Each student also has to write a short essay on the selected topic. Essays must be composed using a given template and must be of length 5-8 pages (including figures, tables, and references). The essay is due in 3 weeks after the presentation. The quality of this essay will be evaluated and considered for the final grade.

It is not sufficient to present the selected papers only, the students should do independent literature research and in their presentations and essays they should summarize the whole topic.

Each student will have a tutor (typically a research assistant from the Advanced Interactive Technologies Group, the Distributed Systems Group, or the Computational Robotics Lab) assigned with whom they can discuss their papers in detail and receive preliminary feedback on their presentation and their essay.

The seminar will be held in English. Presentations and reports must be in English. Attendees are required to participate in all sessions.

Templates

Grading

• the quality of the presentation;
• the quality of the essay;
• participation in discussions and feedback sessions after each presentation.

Schedule

Topics

1) Eye Gaze and Intelligent User Interfaces

Eye tracking is the process of measuring the point of gaze or where the user is looking. This area of research has a long history in medicine and psychology as a tool for understanding human behavior. However, most of the application domains have relied on eye tracking as a passive analysis tool. Recent technological advancements together with affordable eye trackers have created an interest to develop attention and gaze-aware systems. Our eyes can be used as an additional input modality to control user interfaces, help people with disabilities, or create more immersive video games.

Potential papers:

• Toward mobile eye-based human-computer interaction, Perv. Comp. '10
• Gaze+RST: Integrating Gaze and Multitouch for Remote Rotate-Scale-Translate Tasks, CHI '14
• Gaze and Touch Interaction on Tablets, UIST '16
• Pursuits: spontaneous interaction with displays based on smooth pursuit eye movement and moving targets, Ubicomp '13
• The Royal Corgi: Exploring Social Gaze Interaction for Immersive Gameplay, CHI '15
• Everyday Eye Contact Detection Using Unsupervised Gaze Target Discovery, UIST '17

Supervisor: Mihai Bâce

2) Sharing Experiences in Augmented and Virtual Reality Environments

Augmented Reality (AR) and Virtual Reality (VR) create exciting opportunities to engage people in immersive experiences. However, most such applications are designed based on the first-person point of view, of which one drawback is that the experience is user-specific. To solve this problem, immersive collaboration systems are designed to simultaneously involve more people to communicate using video, audio, and gestures. These systems not only tackle practical issues in industry, but also provide a new prototype of social media platform.

Potential papers:

• ShowMe: A Remote Collaboration System that Supports Immersive Gestural Communication, CHI '15
• Hybrid Avatars - Enabling co-Presence in Multiple Realities, Web3D '16
• Webizing Mixed Reality for Cooperative Augmentation of Life Experience, CSCW '16
• Demo: Visualization Technologies for Sharing Visit Experience, SA '16
• Mixed Reality Collaboration through Sharing a Live Panorama, SA '17

Supervisor: Jing Yang

3) Wearables in Healthcare and Wellbeing

Ubiquitous sensor technology and the proliferation of wearable devices to measure physiological data -- such as fitness trackers -- pose new opportunities to management and understand chronic conditions. Through remote and continuous monitoring we can get a better understanding of patients daily life, routine and vital signs. Moreover, the correct interpretation of these data allows space for the development of computational models that would help patients in the self management of their diseases. For example, warning them when their condition is about to worsen, enabling them to take the right measurements.

Potential papers:

• Crafting a View of Self-Tracking Data in the Clinical visit, CHI ’17
• EmotionCheck: Leveraging Bodily signals and False Feedback to Regulate our Emotions, Ubicomp ’16
• Quantifying the Body and Caring for the Mind: Self-tracking in Multiple Sclerosis, CHI ’17
• CrossCheck: Toward passive sensing and detection of mental health changes in people with schizophrenia, Ubicomp ’16
• Smart wearable body sensors for patient self-assessment and monitoring, Archives of Public Health ’14

Supervisor: Liliana Barrios

4) Fast Visual Object Detection

The visual channel is one of the main cues used by humans. In order to create human-computer interaction systems and model the human’s intentions, it is essential to understand what the human sees. A crucial challenge is to detect and recognize the objects in the user’s field of view. Recently developed object detection algorithms based on convolutional neural networks make this possible.

Potential papers:

• SSD: Single Shot MultiBox Detector, ECCV '16
• Mask R-CNN, ICCV ’17
• Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks, NIPS ’15

Supervisor: Vincent Becker

5) Human Motion Prediction

Creating and understanding digital representations of human motion is a long-standing problem in Computer Science and a central aspect in many related research fields such as Computer Graphics, Computer Vision, Robotics, or Biomechanics. Human motion is inherently complex because it is non-linear, high-dimensional and subject to great variability. Furthermore, we as humans are exceptionally skilled at detecting flaws and irregularities in human motion, which sets a high bar for any computational method producing such motion. The suggested papers deal with extracting or generating a human skeleton in time series data using neural networks.

Potential papers:

• VNect: Real-Time 3D Human Pose Estimation with a Single RGB Camera
• Recurrent Network Models for Human Dynamics
• Deep Representation Learning for Human Motion Prediction and Classification

Supervisor: Manuel Kaufmann

6) Human Pose Estimation in the Wild

Estimating human poses is one of the core problems in computer vision and has many applications in life sciences, computer animation and the growing fields of robotics, augmented and virtual reality. Accurate pose estimates can also drastically improve the performance of activity recognition and high-level analysis of videos. Recent pose estimation methods have exploited deep convolutional networks (ConvNets) for body-part detection in single, fully unconstrained images.

Potential papers:

• Single image, single person pose estimation - Convolutional pose machine
• Single image, multiple person pose estimation – top-down
• Single image, multiple person pose estimation – deep cut

Supervisor: Jie Song

7) Policy Search Methods

Policy search methods form a class of algorithms used for finding near optimal control policies for robots, or in general, for solving sequential decision making problems. The standard formulation of the sequential decision making problem assumes discrete action space. This problem can be solved using value function based methods like Q-learning or SARSA. In cases when the action space is continuous, previous methods are hard to apply. For problems with continuous action space, the policy search methods are preferred. In contrast to value function methods which focus on estimating the value function, policy search methods take different approach by searching the policy function directly.

Potential papers:

• A survey on policy search for robotics
• Trust region policy optimization
• Asynchronous methods for deep reinforcement learning

Supervisor: Stefan Stevsic

8) Reinforcement Learning for Discrete Problems

Sequential decision problems are common in robotics and computer vision. Such problems can be tackled by an agent that receives input at each time step and selects an action to perform. A reward signal indicates whether an action was desirable and the goal is to maximize the average reward per step. With accurate models a sequence of desirable actions can be selected with planning methods. However, when no good models are available or when the problem dimensionality is high such planning methods quickly become infeasible. Reinforcement learning deals with learning such agents from simulations when no direct supervision can be given. This topic will include the discussion of POMDP formalism, tabular methods for reinforcement learning (SARSA, Q-Learning, TD-Lambda, Eligibility Traces) and recent methods such as DQN where neural networks are used as function approximators.

Potential papers:

• Reinforcement learning: An introduction
• Playing atari with deep reinforcement learning
• Prioritized experience replay

Supervisor: Benjamin Hepp

9) Sensing in soft robots: methods, systems and implementations

Soft robots have infinite degrees of freedom, and by changing their shape, they can interact either passively or actively with their environment. The change in the shape of a soft robot can be used to infer the external forces generated through passive interactions with the environment or to reason about the active actuation forces required to control its motions. Direct ways of measuring the contact forces between the soft robots and its environment have also been developed. The scope of the current topic is to explore what types of sensing methods are utilized, what kinds of sensing systems have been developed and how they are implemented within the context of soft robotics.

Potential papers:

• Tech Note: Digital Foam, 3D User Interfaces '08
• Batch Fabrication of Customizable Silicone-Textile Composite Capacitive Strain Sensors for Human Motion Tracking, Adv. Mater. Technol., vol. 2, no. 9, '17
• Macrobend optical sensing for pose measurement in soft robot arms, Smart Mater. Struct., vol. 24, no. 12, '15
• 3-Dimensional soft shape sensor based on dual-layer orthogonal fiber Bragg grating mesh, OFC '17
• Embedded electro-conductive yarn for shape sensing of soft robotic manipulators, EMBC '15

Supervisor: Nitish Kumar

10) Learning Approaches for Physically Based Environments

Recent technological advances in general purpose parallel computing put Deep Learning at the core of many fundamental problems in Computer Science. Physics-based animation and robotics are prime examples of areas where promising learning approaching have been proposed to learn physics models, to build fast physics simulation and/or fast decision making algorithms. The scope of this topic is to get familiar with different Machine Learning strategies used in the context of physics simulation and dynamic locomotion for virtual characters.

Potential papers:

• DeepLoco: Dynamic Locomotion Skills Using Hierarchical Deep Reinforcement Learning
• Learning Locomotion Skills Using DeepRL: Does the Choice of Action Space Matter?
• Data-driven Fluid Simulations using Regression Forests
• Data-driven Physics for Human Soft Tissue Animation

Supervisor: Vittorio Megaro

11) Soft Robot Modeling and Control

Soft robots naturally conform to their environment (be this the floor, or a target objects, etc.). This fact makes soft robots inherently safe, and has the promise to enable high performance in the face of environmental uncertainty. However, this fact also makes controlling soft robots very challenging. There is currently no one accepted answer for how to best control even simple soft robots, and state of the art fabricated examples often make use of custom, user-scripted input signals. The scope of this seminar topic is to learn about the different ways researchers have tried to model and control soft robots. Your focus should encompass both control methods that were made specifically for a single given robot, as well as general-purpose control methodologies that could apply to a broad class of soft robots.

Potential papers:

• Control of elastic soft robots based on real-time finite element method
• Design and control of a soft and continuously deformable 2D robotic manipulation system
• Model-free control framework for multi-limb soft robots
• Design, fabrication and control of soft robots

Supervisor: James Bern

12) Dexterous Object Manipulation using Robots

Using manipulator robots for tasks like pick-and-place of rigid objects or assembling of different parts is nowadays state of the art in industry. In order to fulfill more complex tasks, it is becoming a requirement for robots to manipulate objects as dexterously as humans do. To have such capabilities, the robots need to possess a deep understanding of the physical world of deformable systems in order to plan and execute movements in a desired way.

Potential papers:

• Mixed Integer Quadratic Program trajectory generation for a quadrotor with a cable-suspended payload
• Aggressive Quadrotor Flight through Narrow Gaps with Onboard Sensing and Computing using Active Vision
• Dynamic Modeling and Motion Planning for Marionettes: Rigid Bodies Articulated by Massless Strings

Supervisor: Simon Zimmerman

Contact

For further information please contact Mihai Bâce.