available projects

2026

1. Learning robotic juggling

   Valentin Hartmann and Jin Cheng

   dexterous-manipulation, reinforcement-learning, 2026

   Abs PDF

   Dexterous manipulation is an emerging field where more and more capabilities have been shown. These capabilities are usually fairly controlled and slow, and interact with static, or quasi-static objects. In this thesis, we want to learn fast reactive skills, particularly robotic juggling with balls or juggling clubs (as shown on the right side) and a single hand mounted on a UR5. We start out in simulation, and then deploy it on the real robot. Particularly for dealing with juggling clubs, the dexterity of the hand is crucial, compared to juggling with balls, which can be done with simpler end effectors.

2. Learning manipulation from human priors

   Valentin Hartmann and Jin Cheng

   dexterous-manipulation, reinforcement-learning, 2026

   Abs PDF

   Dexterous manipulation is a challenging problem partially due to the high dimensional space of the robot hand, and partially due to the complex robot-object interactions. One example for this is cube reorientation using a dextrous hand. Current approaches to in hand reorientation are usually also relatively imprecise, and do not actually hold the cube at the goal orientation. However, when training this skill from scratch in simulation, relatively conservative behaviours are learned, such as rotating the cube in hand, compared to what a human would do: rotating it by pinch grasping it with the fingertips. More complex, possibly ’more risky’ behaviours are usually not learned, such as pinch grasping the cube and rotating it with other fingers, therefore motivating how we could learn from human data in a manipulation setting, compared to requiring complex reward engineering in order to achieve the desired behaviours. In this work, we want to explore using human motions as prior for learning manipulation behaviours. The cube orientation task is just one of the possible specific tasks that we could tackle. Others include e.g., functional grasping of everyday objects.

3. Benchmarking robotic multi agent reinforcement learning

   Valentin Hartmann

   reinforcement-learning, multi-robot-rl, 2026

   Abs PDF

   In many industrial settings it is desirable to maximise throughput in order to speed up production. One way of doing this is utilizing multiple robots. When dealing with multiple robots, in addition to solving the motion planning problem, we also need to decide which robot does what and at what time. Formulating this as a reinforcement learning problem is relatively straightforward, but depending on the formulation, we need to make sure that the robots stay safe (i.e. do not avoid collisions or actions that would damage the objects we are interacting with). An example for such a multi robot setting is typing with a multi fingered robot hand, or robotic pick and place settings. In settings like these, where we can achieve task success with only a single robot, reinforcement learning approaches tend to find conservative suboptimal policies, where e.g., only a single robot is utilized for solving the task in order to avoid penalties for doing the wrong thing with a robot.

4. Learning Dexterous Robotic Manipulation from Human Demonstrations

   Christoforos Mavrogiannis

   dexterous-manipulation, reinforcement-learning, imitation-learning, 2026

   Abs PDF

   The goal of this project is to endow a robotic arm-hand system with the ability to complete everyday-life tasks like using tools, opening doors, or assembling equipment. This skill will have important implications across a range of domains, including personal robotics, manufacturing, and fulfillment.