Abstract:
One goal of personal robotics is to have robots in homes performing everyday tasks efficiently to improve our quality of life. Towards this end, manipulators are needed which are low cost, safe around humans, and approach human-level dexterity. However, existing off-the-shelf manipulators are expensive both in cost and manufacturing time, difficult to repair, and unsafe to operate with delicate objects due to rigid components. Soft robotic manipulators, on the other hand, show great promise as their compliance allows them to conform to objects, exhibit physical robustness, and execute safe object interactions, while being low cost through the use of rapid prototyping. But designing soft robot manipulators is challenging due to their high degrees of freedom and the inherent complexity of modeling soft materials in simulation. This makes it difficult to iterate on the soft manipulator design prior to manufacturing. To this end, this thesis explores a spectrum of approaches to quickly iterate on the design and evaluation of soft hands, in a matter of days, to allow rapid turnaround of real robot hand prototypes. This allows the designer to swiftly assess the performance of the design for tasks in the real world, and inform future design improvements using real-world measurable metrics. We explore various design iteration approaches by designing and evaluating two different types of manipulators: a parallel delta manipulator and a tendon-driven anthropomorphic (human-like) hand.

Delta manipulators have high precision and low inertia, which lend themselves perfectly to performing fine-grained manipulations. However, 3D-printing these manipulators with soft materials leads to non-ideal kinematic behavior. Our first completed work explores iterating on the design of 3D-printed parallelogram links, which are a key component of the delta manipulator, using human-in-the-loop. We iterate on two parameters, thicknesses of the hinge and the beam of the parallelogram, eventually leading to a design that enables the resulting 3D-printed soft delta manipulator to achieve close to ideal kinematic behavior. Our second completed work explores evaluating a two-fingered 6-DoF delta manipulator, consisting of the designed parallelogram links, using teleoperation in real-world tasks. We demonstrate the compliance and dexterity of our gripper through six dexterous manipulation tasks involving small and delicate objects, such as twisting a grape off a stem. Contrary to first designing the manipulator and then evaluating it in the real world, our third completed work uses a unified design iteration and evaluation framework for a 3D-printed 16-DoF dexterous anthropomorphic soft hand (DASH). We rapidly design and test three iterations of DASH, across a total of 8 days, by leveraging 3D-printing for fabrication and utilizing teleoperation for evaluation across 30 real- world manipulation tasks. Our final iteration of DASH solves 16 of the 30 tasks compared to Allegro, a popular rigid manipulator on the market, which can only solve 7 tasks.

While our completed work uses human-in-the-loop techniques and teleoperation to iterate and evaluate soft manipulator designs, our proposed work explores using automated techniques for design iteration, and using learned policies for evaluation. In our first proposed work, we leverage both human videos and teleoperated demonstrations to learn policies that perform complex tasks with DASH, such as opening a bottle cap. The goal of this work is to demonstrate the capability of DASH on a wider variety of tasks, as compared to our set of 30 teleoperated tasks. Addition- ally, this opens a new modality of evaluation in our framework, besides teleoperation. Our second proposed work focuses on automating the design iteration process, thereby taking the human out of the loop, in order to explore a larger design space. To achieve this, we simulate the soft hand as a rigid body and use hand evolution methods to transfer control policies across a large set of generated hand designs to determine the best design based on the success rate on a pre-specified set of tasks in simulation. Subsequently, we fabricate the design and teleoperate it on real-world tasks to both evaluate its performance, and inform how to change simulation parameters in future iterations. By the end of the thesis, we hope to better understand the strengths and weaknesses of our approaches in terms of effectiveness, the time taken to design a prototype with a set of task goals in mind and to evaluate the prototype in the real world, ease of defining real-world measurable task metrics, and overall turnaround time for the final soft hand design.

Thesis Committee Members:
Nancy Pollard, Co-chair
Jean Oh, Co-chair
Matthew T. Mason
Oliver Brock, TU Berlin

More Information