Presenting ReForm, a Robot Learning Sandbox for Deformable Linear Object Manipulation
Other conference contribution, 2021

Recent success of Deep Reinforcement Learning in games has sparked interest to extend such methods to robotic applications. However, since these require extensive amounts of data to be successful, it is challenging to gather enough real-world experience to make them viable in robotics. In an attempt to solve this problem, several simulation environments have been created to allow robots to learn in simulation, thus reducing the time and cost of such an endeavor. ReForm is our own contribution to the growing number of such resources. At the moment of its conception, there was no other resource which focused on deformable object manipulation tasks. This has recently changed with the release of SoftGym, which includes cloth, rope and liquid simulations and PlasticineLab, which is targeted at Plasticine shaping tasks. ReForm on the other hand, focuses exclusively on Deformable Linear Objects (DLOs), with an emphasis of realistic mechanical properties, ranging from low compression strength ropes, elastoplastic metal cables, to purely elastic rubber bands. The motivation to focus on DLOs comes from their widespread across numerous application areas, such as medical (e.g. suturing), industrial (e.g. electrical wiring) and service robotics (e.g. household cables). ReForm, is intended as a simulation sandbox and a tool for benchmarking manipulation of DLOs, with six distinct customizable environments. The implementation is modular and provides interfaces to change parameters such as end-effector degrees of freedom, type of observation and reward function.

deformable object manipulation

simulation

robot learning

Author

Rita Laezza

Chalmers, Electrical Engineering, Systems and control

Robert Gieselmann

Royal Institute of Technology (KTH)

Florian T. Pokorny

Royal Institute of Technology (KTH)

Yiannis Karayiannidis

Chalmers, Electrical Engineering, Systems and control

2021 International Conference on Robotics and Automation (ICRA 2021)
Xi’an, China,

Subject Categories

Robotics

DOI

10.1109/ICRA48506.2021.9561766

More information

Latest update

7/21/2022