The goal of this thesis is to detect a human pose via camera and apply it to our humanoid robot as part of the “imitate-me” project. Currently the robots hand (pib) can be controlled via a Leapmotion sensor. As a next evolutional step the Leapmotion sensor should be replaced by a camera with intelligent image processing.

This thesis will consist of the following steps:

• Literature research in the field of human pose estimation with reasonable finger joint resolution
• Select an appropriate approach and implement it for pib
• Use the prediction’s output to drive pib into the same pose
• Evaluate how the pose estimation performs for this application

Our goal is to enable machine learning algorithms to learn better control policies. Research has shown that more observations lead to more reliable state estimation. We believe in pressure sensitivity. Therefore we want to add pressure sensitivity to our 3D printable robot’s hand . In order to simplify further steps we suggest to create a visualization tool that is dynamically configurable with respect to the version of the robot, the number and the positions of the pressure sensors. Our initial desired goal is a simple reactive controller that performs certain actions on threshold pressure at certain locations of the hand.

This thesis will consist of the following steps:

• Literature research in the field of measurement technologies and sensor output visualization
• Technology selection and setup
• Implement a visualization tool
• Evaluate the setup

The goal of this thesis is the best possible pose estimation for the “imitate-me” project in the pib universe. As of now, pib has a built-in stereo camera. This thesis is a successor of “Human-Robot Interaction based on Keypoint Detection”. The evaluation target of this thesis might be to demonstrate pib a certain motion that it then is able to recreate from the stereo camera observations.

This thesis will consist of the following steps:

• Literature research in the field of stereo-camera based human gait reconstruction/trajectory reconstruction
• Pib-implementation of stereo keypoint detection and control
• Evaluation of quality metrics for the keypoints

Outsmart theoretical bounds

The goal of these theses is to find one or more approaches to solve arbitrarily sophisticated reinforcement learning problems in the context of humanoid robots. Training a high dimensional complex control policy is hard. We have discovered various promising approaches which we would like to evaluate and potentially integrate. The evaluation goals of the individual theses can be customized. Our ultimate goals are grasping and locomotion.

Some promising approaches:

Multitask learning = multihead optimization: use one “base-model” with multiple decision heads (= at least one fully-connected output layer). All of which are trained at the same time.

Meta learning: create a fast learning base policy.

Transfer learning: Similar approach as multitask learning except you only learn to solve ONE task and then reuse a large portion of the model to learn another one.

Inverse Reinforcement Learning [IRL]: Given the behavior and optionally the dynamics of the system, IRL trains a cost function for actual reinforcement learning. Example reference.

Time variant reward functions: Similar to multitask learning, but this approach learns a trajectory’s components individually without replacing components of the policy model. Example reference.

Goal of this project is to make a first step towards an AR/VR only development environment. In order to accomplish this we want to take the following steps:

• Create vr/ar environment with pib in it (you should be able to move around in the scene)
• Replay trajectories
• Add controls for scene manipulation (we want to modify positions, orientations and add/delete objects within the scene and move the robot’s joints)
• Add physics to the environment
• Create trajectories
• Add trained policy to control pib our humanoid robot, so you can watch it and ideally interact with it.