Fachbereich Informatik

64-476 Oberseminar Technische Aspekte Multimodaler Systeme

Raum F-334
Zeit Dienstag 16:15 - 17:45 Uhr
Veranstalter Jianwei Zhang


18.10. Vorbesprechung und Terminvergabe

25.10. IROS 2016 Highlights and Videos
  Norman Hendrich
Summary: The presentation consists of two parts. I will first summarize my impressions from IROS-2016 with a focus on the Personal Robot Interaction workshop. For the second part, I will select and play a couple of videos from this years' IROS submissions that seem relevant to the TAMS and CML research.
Slides: slides (pdf) 51.0 M pdf

01.11. Multi-Agent-Based Modeling and Simulation in Transport and Logistics
Vortragende: Maryam Reza

15.11. Multimodal localization using WiFi signal strength and 2-D LRF
Vortragende: Benjamin Scholz
Zusammenfassung: In this presentation I explain how Gaussian processes can be used in order to infer a robot's position from Wi-Fi signal strengths. Furthermore I show the implementation of a Wi-Fi position estimation and discuss how it can improve the global localization with a 2D laser range finder.
Folien slides (pdf) 6.0 M pdf

22.11. Policy-based reinforcement learning
Vortragende: Zhen Deng
Zusammenfassung: this talk will focus on policy-based reinforcment learning and some my works based on this methods. and i hope to have a discuss about my future works
Folien slides (pdf) 1.0 M pdf

29.11. Lightning Talks
Vortragende: Michael Görner
Sebastian Starke
Florens Wasserfall
Dennis Krupke
Zusammenfassung: Status updates from several PhD students in the group
Folien slides - Michael Görner (pdf) 0.2 M pdf
Folien slides - Sebastian Starke (pdf) 4.5 M pdf
Folien slides - Dennis Krupke (pdf) 10 M pdf
Folien slides - Florens Wasserfall (pdf) 7 M pdf

06.12. Action Path Improvement in Robotic Rotational Manipulations
Vortragende: Junhu He
Zusammenfassung: This presentation consists of two parts. First I will have a short introduction to robot in-hand manipulation. Besides, a fast review of my previous work will be given. Then I will present how the policy based reinforcement learning methods are used in the robotic rotational manipulations in my research.
Folien Action Path Improvement in Robotic Rotational Manipulations 2.0 M pdf

13.12. Reinforcement Learning for a Self-Improving Walking Pattern with Force Sensors in RoboCup
Vortragender: Fabian Fiedler
Abstract: Humanoid walking of robots has been an active field of research for over 40 Years. One of the currently most widespread approaches for a stable movement is to keep the Zero-Moment-Point (ZMP) inside the support polygon of the stance leg. However calculating the exact position of the ZMP is not possible without deep knowledge of the robot, especially about the actuators. That knowledge covers e.g. the current temperature, voltage and state of the gears of the actuators. This knowledge is often not determinable and therefore it is generally not possible to directly calculate the optimal trajectory with regards to the ZMP.
This talk addresses the current state of the speaker's master thesis to develop integrated force sensors and to use them to determine the ZMP. The thereby found ZMP is used for reinforcement learning to enhance the walking capabilities of the robots used by the department RoboCup team "Hamburg Bit-Bots".
Slides Slides 2.0 M pdf

20.12. Towards an Improved Understanding of Human In-hand Manipulation Capabilities
Vortragende: Eugen Richter
Folien slides 3.8 M pdf


10.01. Online Speech Recognition and Deep Semantics Understanding
Vortragende: Jinpeng Mi
Slides 1.0 MB pdf

17.01. Neural Information Processing Systems 2016 - Barcelona
  Norman Hendrich
Summary: While all papers from NIPS-2016 are open access (see nips.cc), the keynotes and workshop talks are not online. I will summarize the invited talks from Andrew Ng and Yann LeCun, the conference paper review approach ("reduce both bias and variance"), and show selected workshop presentations.
Slides 35.0 MB pdf

24.01. Title
  Marc Bestmann
  canceled. The talk has been postponed to April 2017.

31.01. Real-Time Object Shape Perception via Force/Torque Sensor
  Stephan Rau
Summary: Topic of the talk is the current state of the author's master thesis: Real-Time Object Shape Perception via Force/Torque Sensor.
Modeling of 3D objects in blind scenes can be achieved through haptic exploration. A Force/Torque sensor mounted at the wrist of a robotic arm can be used to obtain force and torque data when an object is touched. The location of the contact is gathered via forward kinematics. The so acquired points of contact are transferred into an Octomap to build up the 3D model of the object. An exploration heuristic has to be applied in order to be able to drive along the surface of the object.
The experimental setup consists of a Mitsubishi PA-10 6-DOF robotic arm, Sunrise M3207 F/T-Sensor and a sphere fixed at the tip of the robotic arm.
Slides 6.6 M pdf