Learning machines: DFKI presents new findings at the Conference on Robot Learning

TacEx - Simulation environment enables dexterous robots

Reaching for a can of soda, picking up a pen, stroking a dog - everyday tasks that we humans don't normally give much thought to. For robots, however, these tasks are complex. Although they have sensors, they do not have a real sense of touch like humans and are not yet able to react well in situations for which they have not been trained. “A robot does not automatically know how to reach for a can of soda without crushing it, because it does not know how much force is needed to grasp the can. This makes their flexible use in unpredictable environments, where they also come into contact with humans, difficult and dangerous”, says Prof Dr Jan Peters, head of the “Systemic AI for Learning Robots” (SAIROL) research department at DFKI. At CoRL, the Darmstadt researchers present TacEx, an innovative solution to this problem. It is a reliable and accurate modular simulator for GelSight tactile sensors. For training robots, simulations are very promising. Through simulated situations in which they solve different tasks, the machines learn for their use in the real world. The innovative framework from Darmstadt enables the use of GelSight Mini sensors for reinforcement learning. As part of the simulation, the robots perform various manipulation tasks and can thus train their fine motor skills.

Step by step: The Potential of Diffusion Policies (DPs) for More Efficient Humanoid Robot Locomotion

In addition to dexterous manipulation, there is another area where robots, unlike humans, require extensive training: their locomotion. Humanoid robots, for example, are promising for use in many fields, but their locomotion is complex due to the few points of contact with the ground.

However, because of their similarity to humans, there is a large amount of data that can be used to train robots. For example, videos of humans performing tasks can be used as a data base. The challenge now is to generate possible rules for the behaviour of robots from this wealth of data in order to train them. Diffusion Policies (DP) have been successfully used in the field of manipulation. “However, reinforcement learning is still widely used for the locomotion of robots. This often results in strange modes of movement,” says Peters. The researchers in Darmstadt have now investigated whether DPs can also be used to train the locomotion of humanoid robots. They found that although robots can develop a stable gait pattern with the help of DPs, successful locomotion training requires a larger and more diverse database than is needed for DPs in the field of manipulation. This starting point enables new approaches to locomotion learning for humanoid robots.

DFKI at CoRL: Paving the way to fully autonomous robots

"An intelligent machine that learns from experience and automatically adapts to its environment is the goal that the CoRL participants are working towards. I am delighted that we are able to present our results in two key areas of robot learning, dexterous manipulation and humanoid robot locomotion, in this innovative environment and contribute to the robotics of tomorrow", says Peters.

The 8th Conference on Robot Learning (CoRL) will take place from 6 to 9 November in Munich. Since its establishment in 2017, CoRL has been dedicated to promoting cutting-edge research and innovation in the field of robotics and machine learning.

DFKI research department head Prof. Dr. Jan Peters is co-organising the LocoLearn workshop and the Morphology-Aware Policy and Design Learning workshop. He will also give a keynote speech at the workshop on Learning Robotic Assembly of Industrial and Everyday Objects.