People are born with the ability to change their behavior according to the objects they hold in their hands and the tasks they are trying to perform. For example, students can learn to carefully remove the outer skin when cutting certain fruits or vegetables, or to cut around tougher parts such as avocados or peach seeds.
Robots must be able to efficiently cut things with mixed material compositions or textures to assist humans with common tasks such as cooking and meal preparation. But transferring this capability to robots has proven very difficult so far.
RoboNinja, a machine learning-based system recently developed by researchers at Columbia University, CMU, UC Berkeley, and other American institutions, can enable robots to cut multi-material objects, especially soft things with hard cores. His articles, published on the ArXiv preprint service, can contribute to improving the skills of robots created to help people with household chores and daily cooking tasks.
Zhenjia Xu, Zhou Xian, and colleagues noted in their paper that RoboNinja aims to remove the soft part of an object while preserving the hard core, thereby maximizing efficiency, unlike previous studies that used open-loop cutting actions to cut single-material objects (such as slicing cucumbers). Our approach uses an interactive state estimator to do this and an adaptive interrupt policy to close the perception-action loop.
Using a computer program, they were able to create a computer program that would allow them to create a computer program that could run on any computer. The goals of the system are to extract as much pulp as possible while reducing collisions with the central seed and using as little force as possible.
Cutting Processes of Robots and Developed Algorithms
According to the paper by Xu, Xian and colleagues, the system uses sparse collision information to repeatedly predict the position and geometry of an object's core, then generates closed-loop interrupt actions based on the predicted state and a tolerance value. According to the statement, “The tolerance value changes the conservatism of the policy when meeting collisions by maintaining an adaptive safety distance from the estimated kernel”.
To evaluate their system for cutting multi-material objects, the researchers developed a cutting simulation environment more suited to the challenge they faced. In this environment, a robot can cut objects made of a combination of hard and soft materials in various ways.
According to Xu, Xian et al.'s paper, “Current simulators are limited in simulating multi-material products or calculating energy use throughout the cutting process. To address this issue, we are creating a differentiable shear simulator that supports multi-material connectivity and enables the creation of sample optimized trajectories for policy learning.
RoboNinja enabled Xu, Xian and colleagues' simulations of the robotic gripper to extract a significant amount of soft material from objects while limiting collisions with hard parts and consuming a tolerable amount of energy. To further validate the framework's performance in real-world scenarios and when cutting objects with various core geometries, the team then tested it on a real robotic gripper.
In their report, the researchers wrote that our trials demonstrated the generalizability of our strategy to innovative core geometries and even to real fruit. “We expect the results of our experiments and the newly created simulator to spur further research on robot learning involving interactions with multi-material elements,” the authors write.
Source: Techxplore
Günceleme: 14/03/2023 14:36