Novel Contact-Aware Robot Design Helps Optimize Complex and Organic Shapes
MIT CSAIL researchers created a new method to computationally optimize the shape and control of a robotic manipulator for a specific task.
Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory have created a new method to represent robotic manipulators, enabling increased biomimicry in future designs. Making machines more like us necessitates a delicate balance between design and control; existing methods for the creation of devices like robotic hands battle trade-offs between design complexities critical for contact-heavy tasks and the practical restraints of manufacturing. Rather than continuing robot manipulation in two stages — design and control — the MIT team proposes joint optimization through the construction of an end-to-end differential framework.
The two key components of their framework are a novel deformation-based parameterization and a differential rigid body simulator. The new method of parameterization allows for the design of articulated rigid robots with arbitrary, complex geometry with a technique called “cage-based deformation.” Essentially, the user is able to change or deform the geometry of a shape in real-time. Meanwhile, the rigid body simulator can handle contact-rich scenarios, computing analytical gradients for a full spectrum of kinematic and dynamic parameters. In testing the framework on multiple manipulation tasks, it outperformed existing methods that only optimize design or control or co-optimize using gradient-free methods.
Their system uses software to manipulate the design and simulate the robot doing a task. It then provides an optimization score to assess the design and control. This task-driven approach to manipulator optimization has the potential for a wide range of applications where tasks need to be performed repeatedly, such as manufacturing and warehouse management. The simulation tools eliminate the need to evaluate the design by manufacturing it and testing it in the real world; they not only find better solutions, but find them faster. The quickly-available optimization scores can significantly shorten the design cycle.
Their paper was presented virtually at the 2021 Robotic Science and Systems conference and details the cage-based algorithm, and the design of a single robotic finger to test it, as well as all experimental data obtained. The team’s plan for future work is to extend the software to optimize the manipulators concurrently for multiple tasks.