Learning Soft Robotic Assembly using Tactile Sensing
We are thrilled to announce that our paper has been accepted to the 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2023) (acceptance rate of 43.3%)!
Joaquín Royo Miquel, Masashi Hamaya, Cristian Beltran-Hernandez, and Kazutoshi Tanaka, “Learning Robotic Assembly by Leveraging Physical Softness and Tactile Sensing” [paper] [project page]
Joaquín Royo Miquel was a research intern at OMRON SINIC X. This work has been done at this internship.
Background
Our objective is to enable robots to complete assembly tasks while reducing the engineering effort to eliminate uncertainties, such as sensor calibrations for the precise location of the pieces or the design of jigs to anchor them.
Several soft robots, composed of compliant materials, have demonstrated adaptability in insertion tasks (see our previous blog posts [Hamaya+, 2020] [Tanaka+, 2021]).
However, additional sources of uncertainty, such as variations in the grasping angle, may lead robots to fail the task, for instance, by initiating insertion without properly positioning the part.
This study integrates a tactile sensor into a soft robot toward a more robust assembly. Tactile signal patterns can mark the timing for subtask transitions even in scenarios with imprecise goal positions and grasping poses. In the context of the peg-in-hole task, it is assumed that the presence of a hole can be detected by identifying the distinctive dynamic pattern that the peg generates on the tactile sensor when it encounters the edge of the hole. This is framed as an anomaly detection task.
To accomplish the insertion task, a model-based online learning controller guides the peg’s tip toward the hole. We also incorporate a pose estimation method with a force-torque sensor.
Overview of Our Proposed Method
Fig. 1 shows the overview of the proposed method composed of (a) a tactile-based hole detection for subtask transition and (b) a search and insertion controller.
The hole detection is used to distinguish the tactile signal patterns generated upon peg-hole contact, thereby triggering the insertion. This study used a vision-based tactile sensor, Digit [Lambeta+, 2020]. We perform signal pre-processing on the tactile sensors to extract frequency features and then apply anomaly detection composed of a Variational AutoEncoder (VAE). The assumption is that the frequency features
undergo significant changes when the peg’s tip fits the hole,
and the anomaly detection method can inform these changes.
The search controller is used to navigate the tip of the peg to the hole. The controller is obtained by a model-based learning approach. We train the forward model of the robot and select the optimal action using a Cross-Entropy Method. Once the cue is received, the search controller is switched for the alignment and insertion controller. Peg tip pose is estimated using tactile and force-torque sensors.
Experiments
We performed real robot experiments using a soft wrist [Tanaka+, 2020] and a tactile sensor.
The results demonstrate that our method achieves a 100\% success rate in scenarios with less uncertain goal pose ( σ= 2mm) and grasp misalignment (up to 5°) and a 70% success rate in scenarios with uncertain goal pose (σ = 10mm) and grasp misalignment (up to 20°) shown in Fig.2. Moreover, our anomaly detection model can generalize to different peg diameters without additional training.
Future works
We will explore faster task completion and extend our method to adapt to various part shapes.
If you are interested in our internship, please check our website!
If you are interested in collaborating with us, please send us an e-mail! contact@sinicx.com