Site Search  

Research » Research Projects


For updated information, you may visit our new website:

https://am.is.tuebingen.mpg.de



Learning Locomotion with the Little Dog quadruped robot: This project explores learning autonomous quadruped locomotion over very rugged and unknown terrain. Read more ...

Autonomous Robotic Manipulation: This project explores grasping and manipulation of unknown objects in a large variety of scenarios, from simple reach and grasp to exploring objects without vision in a gym bag. Read more ...

Associative Skill Memories: Uncertainty in the sensory- motor system as well as uncertainty in real world en- vironments often times leads to imprecise reaching and grasping behavior. Our approach aims to acquire a memory- base predictive model for each sensor using knowledge from previous similar task executions. Read more ...

Complaint Manipulation with Force Feedback: Personal robots can only be- come widespread if they are capable of safely operat- ing among humans. In uncertain and highly dynamic en- vironments such as human households, robots need to be able to instantly adapt their behavior to unforseen events. The main idea of our approach is to online modify the desired trajectory generated by the DMP us- ing sensor information from previous task executions. Read more ...

Template-Based Learning of Model Free Grasping: The ability to grasp unknown objects is an important skill for personal robots. Our approach is based on the simple as- sumption that similar objects can be grasped with similar grasp configurations. Read more ...

Probabilistic Object Tracking using a Depth Camera: The problem we are trying to solve is to track an object while it is being manipulated by a robot. In order to acquire images we use a depth camera, such as the Microsoft Kinect. The core prob- lem consists in determining how the object has moved between two instants in time, t0 and t1, given the depth images acquired at these instants. Read more ...

Model-based control of floating-based robots: Legged robots that can loco- mote on very rough terrain and perform agile and dexter- ous tasks need high-performance controllers. In contrast to very stiff position controlled robots, modern legged platforms have torque control capabilities which offer interesting possibilities for high performance controllers that both guarantee good tracking performance (e.g. for precise foot placement) and compliance with the envi- ronment (e.g. to absorb unexpected disturbances). Read more ...

Stochastic Trajectory Optimization for Motion Planning: A motion planner is a key component in a complete robotics system. It allows the robot to synthesize new motions in situations that it has not previously experienced. We propose a motion planning algorithm based on stochastic trajectory opti- mization that (1) intrinsically generates smooth trajecto- ries, and (2) optimizes arbitrary trajectory costs, such as collisions, motor torques, and task constraints. Read more ...

Human Reinforcement Learning uses Reward Weighted Averaging: Many human movement skills require optimizing a movement such that a future event has a desired outcome. Such skills are, e.g., hitting a ball with a bat or swinging a golf club to achieve that the ball reaches a particular target. In such tasks, often the main feedback is success or failure, and no clear information is provided how to improve, as, in con- trast, when explicit error feedback is provided and a gradient how to improve. In this project, we investigated this ques- tion using a behavioral paradigm mimicking a ball- hitting task. Read more ...

Local Regression in Dual Space Formulation: For high dimensional move- ment systems like humanoid robots accurate analytical internal models are often not readily available. Thus, an essential component for motor control of intelligent sys- tems is the learning of these internal models. Approxi- mating control models is a nonlinear function problem, which is the focus of this research project. Read more ...

Fusing Visual and Tactile Data for 3D Scene Understanding: To successfully act in an envi- ronment, a robot needs to form an understanding of the geometric structure of this environment. This includes for example the global shape of an object to do better grasp planning or finding free space on a table to place an already grasped object. In this project, we consider the problem of predict- ing complete scene structure from only partial information. Read more ...

3D Object Reconstruction by Fusing Visual and Tactile Data: In this project, we consider the problem of object shape estimation when the state of the world is only partially observable. Read more ...

Pre-Grasp Manipulation: Based on human studies we propose a data-driven framework which is designed to automatically perform pre-grasp manipulation actions to fetch an object from a surface. Read more ...

Legged Locomotion on Difficult Terrain: A recent project started at the AMD concerns biped locomotion on difficult terrains. We are interested in developing a complete system for locomotion in challenging terrains such as the ones found on a disaster site. Read more ...
Designed by: Nerses Ohanyan & Jan Peters
Page last modified on June 06, 2016, at 12:34 PM