Learning Fast and Precise Pixel-to-Torque Control: A Platform for Reproducible Research of Learning on Hardware


Steffen Bleher, Steve Heim, Sebastian Trimpe

  The four steps of the pipeline can be completed in fewer than 18h, only 13 h of hardware interaction time, which is largely automated.


In the field, robots often need to operate in unknown and unstructured environments, where accurate sensing and state estimation (SE) become a major challenge. Cameras have been used with great success in mapping and planning in such environments as well as complex but quasi-static tasks, such as grasping, but are rarely integrated into the control loop for unstable systems. Learning pixel-to-torque control promises to enable robots to flexibly handle a wider variety of tasks. While reinforcement learning (RL) offers a solution in principle, learning pixel-to- torque control for unstable systems that require precise and high-bandwidth control still presents a significant practical challenge, and best practices have not been established. Part of the reason is that many of the most auspicious tools, such as deep neural networks (DNNs), are opaque: the cause for success on one system is difficult to interpret and generalize.

A video synopsis can be found at https://youtu.be/S2llScfG-8E and in the supplementary material (available at https://doi.org/10.1109/MRA.2021.3129448).


IEEE Xplore