Bimanual Motion Primitives for Advanced Planar Robotic Manipulation of 3D Deformable Objects with non-actuated End-effectors

Abstract

Deformable Object Manipulation is a challenging task in Robotics and Automation, as robots need specialized equipment or manual intervention to handle such objects. Deformable objects are complex networks of smaller structures, which makes it challenging to represent their physical relations accurately. Predicting their responses to external stress is also complicated due to the complexity of their internal interactions and varying physical parameters. As these objects have their dimensional values within the same range, their differences are neglectable, allowing for myriad ways they can be deformed in 3D space. Nonetheless, modeling such objects becomes more complex as it involves additional constraints and degrees of freedom. Additionally, capturing the entirety of a 3D deformable object from a single viewpoint can be challenging. As a result, observation and manipulation from various angles are necessary, further complicating the process. All of these factors contribute to the significant increase in computational complexity involved in three-dimensional deformable object manipulation.

Addressing the scenario of food industrial manufacturing lines processing three-dimensional deformable objects like meat or fruits in containers like boxes, this research aims to utilize recent advances in robot manipulation to solely handling the products without supports from specialized devices. Dual-arm robots are utilized here, as shown in many prior studies in deformable object manipulation, they have increased flexibility and capability in handling and controlling soft products. To minimize food hazards during the operation, actuated end-effectors such as grippers are replaced with non-actuated ones like chopsticks. Therefore, the research goal is set as the generation of bimanual motion primitives for advanced robotic planar manipulation of three-dimensional deformable objects.

To address the issue, an estimation on object volumes that acts like a rigid object is proposed to identify the object part that can be used for robotic motion planning. This approach only requires geometric data from a stereo-vision system. Prior trial and error is used to reduce uncertainties caused by unknown object deformability. Based on the estimation, the planar trajectory for translating and rotating goal are formulated as a sequence of deforming and sliding motions. The Intersection-over-Union evaluation metric is used to assess the precision of the proposed estimation-based manipulation. The data obtained from the analysis offer insights into the reliability and opportunities of the approach.