This work proposes a novel framework for insertion-type tasks with soft bodies, such as cleaning a bottle with a soft brush (using soft tools to explore complex paths)
First, a multimodal model based on vision and force perception is trained,
where we use domain randomization to ease the simulation-to-real difficulty of models with soft bodies.
Second, we propose a dynamic safety lock method based on force perception,
which is embedded in the training model to improve the success rate (making the tool explore and traverse the hole's path successfully)
and the security (tools and holes are not damaged) of tasks.
Finally, we perform experiments in simulation and the real world, and the success rate of our proposed method reaches 85.14% in simulation and 83.45% in the real world.
Ablation experiments in the real world demonstrate that our method is effective for complex paths and soft bodies with varying deformation intensities.
(This work is an extension of the ICRA workshop work.
The most significant difference is that this work focuses on the robot's ability to insert soft bodies and explore complex paths,
while the basic workshop work focuses on the shape matching capabilities of rigid body insertions.)
The overview of the robotics soft bodies insertion policy. We use the MUJOCO as the simulator to build up the whole sim-environment, which contains the hole, soft objects and manipulator. The RL model provides the policy to the robot, which contians the domain randomization method. We test different holes and soft bodies in the real world to demonstrate the effectiveness of the RL model.
Network structure of robot insertion models. The entire network is divided into three parts, the environment, the agent network, and the dynamic safety lock (DSL). DSL dynamically compensates the movement of the end of the robot, and the degree of compensation depends on the magnitude of the force feedback and the magnitude of the gain.
Trainning performance. VFTM means the model combine the F/T sensor's information and the visual informantion, we compare three types of model. (VFTM vs. VFTM with admittance controller vs. VFTM with DSL)
Perform domain randomization on the observation at hole locations. We use the same hole and the same object to test.
Perform domain randomization on the soft bodies, which contains the soft body's stiffness and damping. We test the soft bodies with different deformation intensities.
You can see here how different holes are inserted with different objects.
@inproceedings{liu2023a,
title ={A Robust and Safe Strategy for Robotic Assembly},
author ={Yi Liu, Andreas Verleysen and Francis wyffels},
booktitle ={Embracing Contacts - Workshop at ICRA 2023},
year ={2023},
url ={https://openreview.net/forum?id=CN2Zr3ZYHF}
}