Motivation:
The cyber-physical systems group developed a system that allows verifiable safe robot motion planning in the presence of humans. We can guarantee that there will be no safety-critical collisions between the robot and the human.
Most recently, this system has been used to show that it is possible to perform motion planning with deep reinforcement learning techniques while still ensuring absolute safety of all humans in the scene.
This is an important step towards a safe and intelligent human-robot interaction. However, this system was mainly developed in simulation. Therefore, the goal of this course is to develop a more realistic reinforcement learning setup.
This laboratory is a hybrid of theoretically conveyed contents, which are applied practically and further developed in subsequent group work.
The main practical development areas will be:
- Human motion analysis: Right now, we work with a very limited set of human motion captures provided by CMU. As you might already know, for deep learning applications it is crucial to cover as many real-world situations as possible to guarantee universal applicability. Therefore, your goal is to capture new real-world motions, develop algorithms to artificially generate human motions and finally train and improve the RL agent based on your newly created dataset.
- Sim2Real: Your goals will be to set up the communication between the RL pipeline and a real robot, develop methods to guarantee that an RL agent trained in simulation also works in real life, and try to train an RL agent on a real robot.
- Improving the agent: There are many possible ways of improving the existing RL agent. This could include an extensive study of the RL algorithm and its parameters, development of new training methods, or improving simulation speed.
The cyber-physical systems group developed a system that allows verifiable safe robot motion planning in the presence of humans. We can guarantee that there will be no safety-critical collisions between the robot and the human.
Most recently, this system has been used to show that it is possible to perform motion planning with deep reinforcement learning techniques while still ensuring absolute safety of all humans in the scene.
This is an important step towards a safe and intelligent human-robot interaction. However, this system was mainly developed in simulation. Therefore, the goal of this course is to develop a more realistic reinforcement learning setup.
This laboratory is a hybrid of theoretically conveyed contents, which are applied practically and further developed in subsequent group work.
The main practical development areas will be:
- Human motion analysis: Right now, we work with a very limited set of human motion captures provided by CMU. As you might already know, for deep learning applications it is crucial to cover as many real-world situations as possible to guarantee universal applicability. Therefore, your goal is to capture new real-world motions, develop algorithms to artificially generate human motions and finally train and improve the RL agent based on your newly created dataset.
- Sim2Real: Your goals will be to set up the communication between the RL pipeline and a real robot, develop methods to guarantee that an RL agent trained in simulation also works in real life, and try to train an RL agent on a real robot.
- Improving the agent: There are many possible ways of improving the existing RL agent. This could include an extensive study of the RL algorithm and its parameters, development of new training methods, or improving simulation speed.
- Dozent: Matthias Althoff
- Dozent: Jakob Thumm