Periodic Reporting for period 4 - 3DPBio (Computational Models of Motion for Fabrication-aware Design of Bioinspired Systems)
Berichtszeitraum: 2024-08-01 bis 2025-01-31
Most robots today are designed to be either as rigid as possible (e.g. industrial Kuka robot arms, Boston Dynamics' Spot and Atlas, etc), or they are fully soft (e.g. Soft Robotics' mGrip). In contrast, the morphological designs we see in nature are fundamentally different. Vertebrates, for example, feature a rigid skeletal system that is complemented by various types of soft tissues (articular cartilage and ligaments, muscles and tendons, soft pads that modulate the forces arising from physical interactions with external objects, etc). Indeed, structural compliance arising from the strategic use of soft materials is an integral part of the design of every biomechanical system, defining the performance, efficiency and robustness of its movements. In our quest to create robots that are as skilled and adaptable as their living counterparts, developing designs that seamlessly integrate soft and rigid materials is therefore of utmost importance. Nevertheless, despite decades of research, such hybrid soft-rigid designs are still out of reach.
In defining how future generations of robots will be made, additive manufacturing (AM) technologies will play a pivotal role. This is because they allow us to create designs of unparalleled geometric complexity using a constantly expanding range of materials. Indeed, if past developments are an indication, within the next decade we will be able to fabricate physical structures that approach, at least at the macro scale, the functional sophistication of their biological counterparts. However, while this unprecedented capability enables fascinating opportunities, it also leads to an explosion in the dimensionality of the space that must be explored during the design process. As AM technologies keep evolving, the gap between "what we can produce" and "what we can design" is therefore rapidly growing.
To effectively leverage the extraordinary design possibilities enabled by AM, our project aimed to develop the computational and mathematical foundations required to study a fundamental scientific question: how are physical deformations, mechanical movements and overall functional capabilities governed by geometric shape features, material compositions and the design of compliant actuation systems? By enabling computers to reason about this question, our work has established new ways to algorithmically create digital designs that can be turned into mechanical lifeforms at the push of a button.
In defining how future generations of robots will be made, additive manufacturing (AM) technologies will play a pivotal role. This is because they allow us to create designs of unparalleled geometric complexity using a constantly expanding range of materials. Indeed, if past developments are an indication, within the next decade we will be able to fabricate physical structures that approach, at least at the macro scale, the functional sophistication of their biological counterparts. However, while this unprecedented capability enables fascinating opportunities, it also leads to an explosion in the dimensionality of the space that must be explored during the design process. As AM technologies keep evolving, the gap between "what we can produce" and "what we can design" is therefore rapidly growing.
To effectively leverage the extraordinary design possibilities enabled by AM, our project aimed to develop the computational and mathematical foundations required to study a fundamental scientific question: how are physical deformations, mechanical movements and overall functional capabilities governed by geometric shape features, material compositions and the design of compliant actuation systems? By enabling computers to reason about this question, our work has established new ways to algorithmically create digital designs that can be turned into mechanical lifeforms at the push of a button.
The research accomplished in the context of this project has been organized into three complementary lines of investigation. The main objectives of these research streams, along with the goals that were achieved, are summarized below. We have disseminated our results through publications in top-tier journals and presentations at leading conferences in visual computing and robotics.
Stream 1: Robotic Materials
The goal of this research stream was to develop computational tools for physics-based modeling and optimization of 3D-printable structures designed to undergo large deformations.
A particular focus was placed on two-dimensional sheet materials that are both strong and flexible, thereby providing robots with tailored protection and actuation capabilities. For example, we developed methods for designing 3D-printable sheet materials with locally controllable properties, enabling the construction of highly customized soft robots and actuators. Our research on scaled sheet materials paves the way for robotic skins that strike an ideal balance between protection and flexibility. Along similar lines, we formalized and automated the design of generalized chainmail materials that uniquely combine strength and flexibility. In the broader context of three-dimensional materials, we developed a novel approach for designing large-scale cellular materials with tailored mechanical properties using differentiable Voronoi diagrams.
We also explored the potential of machine learning techniques for material characterization and design. Our main contributions include a new topology optimization framework that leverages neural implicit representations for the automated synthesis of highly optimized 3D-printable materials. Additionally, we developed a neural formulation for learning mechanical behavior across large and complex material spaces, opening new avenues for the inverse design of microstructures with targeted behaviors.
Stream 2: Computational Methods for Mechatronic Systems
This research stream focused on developing novel algorithms for the co-design of motions, actuator configurations, and force transmission elements—such as rigid or flexible linkages—for new types of mechatronic systems.
Our work in this area led to the development of a first-of-its-kind computational design system for complex animatronic mechanisms, as well as numerical optimization methods for multi-degree-of-freedom mechanisms that incorporate both rigid and compliant components. We also explored the design of mechanical systems that exploit compliance to generate nonlinear periodic motions, using a novel Harmonic Balance approach. Furthemore, we introduced a new method for control-aware design optimization of quadrupedal robots, integrating morphology and locomotion capabilities.
Stream 3: Motion Controllers for Hybrid Soft-Rigid Robots
The goal of this research stream was to formalize the process of developing control policies for hybrid rigid-soft robots. We focused in particular on bio-inspired motion controllers for agile locomotion and manipulation behaviors involving rich interactions with the environment.
As a fundamental building block, we developed a new type of differentiable simulator to facilitate design tasks related to the structure and control of hybrid robots. For instance, we used this differentiable simulator to generate more lifelike motion skills for quadrupedal robots using real-world data. Other key achievements include the development of a novel learning-based control framework that incorporates keyframing to encode high-level locomotion objectives, as well as new methods that combine model-based optimal control with reinforcement learning to achieve versatile and robust legged locomotion.
Stream 1: Robotic Materials
The goal of this research stream was to develop computational tools for physics-based modeling and optimization of 3D-printable structures designed to undergo large deformations.
A particular focus was placed on two-dimensional sheet materials that are both strong and flexible, thereby providing robots with tailored protection and actuation capabilities. For example, we developed methods for designing 3D-printable sheet materials with locally controllable properties, enabling the construction of highly customized soft robots and actuators. Our research on scaled sheet materials paves the way for robotic skins that strike an ideal balance between protection and flexibility. Along similar lines, we formalized and automated the design of generalized chainmail materials that uniquely combine strength and flexibility. In the broader context of three-dimensional materials, we developed a novel approach for designing large-scale cellular materials with tailored mechanical properties using differentiable Voronoi diagrams.
We also explored the potential of machine learning techniques for material characterization and design. Our main contributions include a new topology optimization framework that leverages neural implicit representations for the automated synthesis of highly optimized 3D-printable materials. Additionally, we developed a neural formulation for learning mechanical behavior across large and complex material spaces, opening new avenues for the inverse design of microstructures with targeted behaviors.
Stream 2: Computational Methods for Mechatronic Systems
This research stream focused on developing novel algorithms for the co-design of motions, actuator configurations, and force transmission elements—such as rigid or flexible linkages—for new types of mechatronic systems.
Our work in this area led to the development of a first-of-its-kind computational design system for complex animatronic mechanisms, as well as numerical optimization methods for multi-degree-of-freedom mechanisms that incorporate both rigid and compliant components. We also explored the design of mechanical systems that exploit compliance to generate nonlinear periodic motions, using a novel Harmonic Balance approach. Furthemore, we introduced a new method for control-aware design optimization of quadrupedal robots, integrating morphology and locomotion capabilities.
Stream 3: Motion Controllers for Hybrid Soft-Rigid Robots
The goal of this research stream was to formalize the process of developing control policies for hybrid rigid-soft robots. We focused in particular on bio-inspired motion controllers for agile locomotion and manipulation behaviors involving rich interactions with the environment.
As a fundamental building block, we developed a new type of differentiable simulator to facilitate design tasks related to the structure and control of hybrid robots. For instance, we used this differentiable simulator to generate more lifelike motion skills for quadrupedal robots using real-world data. Other key achievements include the development of a novel learning-based control framework that incorporates keyframing to encode high-level locomotion objectives, as well as new methods that combine model-based optimal control with reinforcement learning to achieve versatile and robust legged locomotion.
Our investigations into the three research streams outlined above have led to more than 30 peer reviewed articles which have been published in the top venues for robotics and visual computing research. Through these publications, we have been able to push forward the state of the art in the areas of robotic materials, computational design of hybrid soft-rigid robotic systems, and motion control for bio-inspired robots, as summarized above.