Robotics, Machine Learning,
Mail. Fabio Muratore
TU Darmstadt, Fachgebiet IAS
Honda Research Institute Europe
63073 Offenbach am Main
Office. Room E323,
There is a large consent both in academia as well as in industry that physical human-robot interaction is attributed a large potential for future robotic applications. While being a generic technology, applications emerge increasingly in the factory domain, particular in the production and assembly processes.
Learning concepts for manipulation tasks is, however still rather academic as it is, imposing a number of assumptions on the underlying problem, and requiring scientists to produce the results. In particular, learning of force-based manipulation is mainly realized using kinesthetic teaching, necessitating expensive and specialized hardware as well as an expert.
The joint research project Motor Dreaming between IAS and HRI takes a different perspective on the problem. It targets to combine data driven learning and exploitative learning in an efficient way. As an alternative to building a skill representation exclusively from data, the core idea is to make additional use of generative models that allow an internal simulation of the task, also known as mental rehearsal. This step involves devising physical simulation models from the real situation, and being able to play them through in different variations. Such a mental rehearsal allows to incorporate uncertainty, such aiming to increase the robustness of reproduction by learning solutions that can deal with large parameter variations.
Chek out my open source repository for reinforcement learning from randomized simulations SimRLacra.
I worte a very easy to follow blog post on quantifying the transferability of sim-to-real control policies at sim2realai.github.io.
The subsequent videos are part of the supplementaty material to our 2018 CoRL paper and show comparison of SPOTA, EPOpt, TRPO and LQR policies on the ball-on-plate task.
Spotlight talk at CoRL 2018
Evaluation of SPOTA, EPOpt, TRPO, and LQR policies in Vortex varying selected physics parameters of the simulation
Cross-evaluation of SPOTA, EPOpt, TRPO, and LQR policies trained in Vortex and in Bullet then tested in both
Appraoch me any time to suggest your ideas
|Year||Type||In coorperation with||Student(s)||Topic|
|Year||Type||In coorperation with||Student(s)||Topic||Document|
|2019||Student Project||GCC||6 groups à 2 students||Parallelizing a Reinforcement Learning Algorithm|
|2019||Bachelor's Thesis||Benedict Flade & HRI||Simon Kohaut||Simultaneous Map Correction and Auto Calibration for Hybrid Localization|
|2019||Bachelor's Thesis||HRI||Robin Menzenbach||Benchmarking Sim-2-Real Algorithms on Real-World Platforms|
|2019||Bachelor's Thesis||Boris Belousov & HRI||Christian Eilers||Bayesian Optimization for Learning from Randomized Simulations|
|2019||Seminar||--||3 groups à 3 students||Reinforcement Learning Class|
|2019||Integrated Project||Boris Belousov||Jonas Eschmann, Robin Menzenbach, Christian Eilers||Underactuated Trajectory-Tracking Control for Long-Exposure Photography (part II)|
|2019||Master's Thesis||HRI||Markus Lamprecht||Benchmarking Robust Control against Reinforcement Learning Methods on a Robotic Balancing Problem|
|2018||Integrated Project||Boris Belousov||Jonas Eschmann, Robin Menzenbach, Christian Eilers||Underactuated Trajectory-Tracking Control for Long-Exposure Photography (part I)|
|2018||Master's Thesis||HRI||Felix Treede||Learning Robust Control Policies from Simulations with Perturbed Parameters|
Robotics, Physics Simulations, Machine Learning, Automatic Control