SSD Seminar Series with Emma Cramer, M.Sc.

Monday, December 11, 2023, 4 to 5pm

On December 11, 2023, we welcome Emma Cramer, M.Sc., of the Chair of Data Science in Mechanical Engineering at RWTH Aachen University as speaker in our SSD Seminar Series. 

The Seminar will take place in Rogowski Building, Room 115, Schinkelstraße 2, 52062 Aachen.


The best of both worlds – combining classic control and reinforcement learning for contact-rich assembly robots

picture of Emma Cramer, M.Sc. Copyright: © Emma Cramer, M.Sc.

Adapting to changing environments or tasks remains a challenge for automatic assembly robots relying on classical control strategies. In recent years, learning-based methods such as Reinforcement Learning (RL) have shown promising results in designing flexible controllers for contact-rich assembly tasks. However, relying on only RL alone tends to be data inefficient, unstable, and difficult to train. One way to mitigate this is to use residual reinforcement learning, which combines classic control algorithms with a reinforcement learning policy. This method thrives in complex robotic manipulation tasks where good but imperfect controllers are available. We investigate a polishing task wherein a robotic arm is tasked with sweeping across various trajectories. This task is a typical example of a manipulation skills involving contact and friction which are inherent to many robotics assembly tasks.

With this, we aim to address several questions:

  1. Can this hybrid approach outperform traditional model-free RL methods in terms of sample efficiency?

  2. How effectively does this method facilitate zero-shot generalization?

  3. How can we leverage the uncertainty inherent in the RL critic to balance the influence of control and RL policy?