Generative Design as design automation
In the engineering world, ‘generative design’ is a term that is constantly up for debate. Some think it’s all about churning out a smorgasbord of ‘solutions’ for an engineer to down select from to meet some multi-objective criteria. Others believe the term can also apply if the algorithm spits out a single, golden solution that checks all the boxes, that might be generative design too?
And let’s not even wade into the murky waters of those who conflate the ‘Generative AI’ label on anything that smells like a simulation-driven algorithm. We’ll save that another day.
Amidst this ongoing debate, a recent paper from researchers at the University of Washington entitled Computational Design of Passive Grippers explores the generative design of passive robotic grippers that have no additional actuation and instead leverage the existing degrees of freedom in a robotic arm to perform grasping tasks.
We have seen this approach taken to automatically ‘generate’ manufacturing aids such as jigs and fixtures in the past, but the researchers must also take into account a more complex automated insert path and stability of the part in motion.
To achieve this, they addressed the key challenge of jointly optimizing the shape and the insert trajectory to ensure a passively stable grasp.
automation without actuation
The paper addresses the challenges of automatically optimizing passive grippers for specific tasks.
The optimization process involves a large-scale search, which is computationally expensive due to the need for physical evaluation and nested optimization.
The researchers introduce two fundamental approaches to handle this complexity.
First, they identify that stability is a function of the contact points between the gripper and the object, termed the ‘grasp configuration’ (GC).
Second, they propose a co-optimization approach that abstracts the gripper geometry into a parametric skeleton, simplifying the search for a feasible insert trajectory.
The researchers then employ a ranking strategy for GC candidates, which serves as the basis for co-optimizing the trajectory and gripper abstraction. The method is validated through an experimental dataset, demonstrating its efficacy in handling a variety of objects with high grasp reliability and stability.
Previous research in the realm of generative gripper design has predominantly concentrated on active grippers. Early computational design efforts targeted antipodal grasping, employing both formulaic approaches and neural network techniques to develop shaped fingertips. Another avenue of research has been vacuum-based gripper design, where a 3D printable manifold and superstructure are generated based on user-specified target locations. The paper at hand distinguishes itself by extending these methodologies to the design of passive grippers, thereby filling a gap in the existing literature.
Grasp Configuration Generation
The paper identifies the concept of Grasp Configuration (GC) as a pivotal element in the optimization of passive grippers. GC refers to the set of contact points between the gripper and the object, and it serves as a design variable that should be optimized.
The researchers propose a co-optimization strategy that involves abstracting the gripper geometry into a parametric skeleton. This abstraction simplifies the search for a feasible insert trajectory by connecting the points on the GC to the flange frame’s origin (FFO) of the robot.
The paper argues that the feasibility of finding a successful insert trajectory is intrinsically linked to the choice of GC. Therefore, the researchers introduce a strategy for computing a list of stable GC candidates that are likely to enable a collision-free insert trajectory.
The researcher then used a discrete topology optimization to generate the gripper design.
This optimization is performed over a collision-free volume, which is determined based on the known trajectory of the object moving away from the gripper. For boundary conditions, forces exerted by the known grasp configuration are set along the normal directions, and certain parts of the gripper around the flange frame’s origin are fixed.
Post-processing involves the application of a smoothing kernel and the use of marching cubes to produce a smooth mesh.
The gripper geometry is further refined by adding small spheres at each contact point, thereby increasing the area of contact and enhancing the robustness of the grasp.
Finally, robot-specific mounting structures, such as holes and a mounting plate, are added for quick installation.
The researchers also make the software tools used for these processes available on GITHUB.
Of the 23 experiments, 21 lead to successful pickups, 17 of these experiments had 100% success rate and only two had success rates below 80% show that our method is able
to successfully automate the design process and displays high grasp reliability in real world experiments.
You can read the full paper by Milin Kodnongbua, Ian Good, Yu Lou, Jeffrey Lipton, Adriana Schulz of University of Washington, ACM Transactions on Graphics at the proceedings of SIGGRAPH 2022 and check out the video of the presentation of their research.
The study does acknowledge several avenues for future work including the stability evaluation for selecting grasp configurations, which could make the method more robust to real-world variables like object placement and gripper trajectory.
The paper also suggests that future research could explore different post-grasp trajectories and consider external forces that may occur during pickup as well as investigating the optimization techniques to handle multiple input shapes or classes of shapes, thereby broadening the applicability of their algorithm.
Thanks to all of the researchers for further enabling the AI driven robot uprising.