Key Take-aways

Key take-aways from “How to Make Better Soft Robots Through Design Optimization”

• Generative design is undervalued and can be a powerful tool

  • Using AI to ‘brainstorm’ a library of feasible designs

  • Importance of graphical tools for design and Ashby plots

• Co-design of robot control and behavioral policy

  • Does co-optimization entail inherent performance compromises?

• Significance of finding minimal models that capture behavior

  • We always need to generate a computational model for material

  • When is simple too simple?

  • Open challenge: considering dynamics in design problems (without extensive reliance on computation)

  • For inflatables, we could study inflation in low Reynolds number regimes (i.e. use viscous fluids)

• Exploit symmetries in systems to reduce modeling complexity

• The trade-off between high fidelity material models and computation requirements

• Soft materials simulation

• Improve sim2real by:

  • Domain randomization and training across multiple situations

  • Injecting noise into the simulator to develop robust policies

  • Bringing in first-principles

  • Knowing the limitations of black-box algorithms

  • Speeding up simulations

• There are opportunities for both qualitative and quantitative methods in inverse design

  • Humans in the loop model for design: how do we quantify this?

• Thinking beyond the ‘unit cell’ when optimizing the topology

  • Global vs. local deformation

• Harnessing traditionally ‘undesirable’ mechanical phenomenon, like buckling instabilities

• Search for general design rules in mechanical systems

• We need a more well-defined framework for morphological computation and embodied intelligence

  • Not just to observe the emergent behavior in the substrate but also to design for the desired behavior

• Integration

  • We can design robot building blocks, but can we unite them and expect the same behavior?

  • Are there certain building blocks where the behavior as a sum is more than its building blocks?

  • With a standardized set of soft robotic parts, can we systematize soft robot designs?

• What are we missing in robotics in general

  • Benchmarks

  • Datasets

  • Especially in the space of computational design: what are the most common actuators/sensors, what are the most common testing protocols, what are the performance measures?

  • Is there a more generic performance measure instead of being outcome-oriented (for specific tasks like grasping, locomotion, ….)

• Standards in Soft robotics

  • Agree on common benchmarks for things such as:

    • Actuation (blocked force), locomotion (speed, cost of transport), shape matching (Hausdorff distance)

  • Fabrication, testing, and measurement standards

  • Make your paper’s data open-source!

  • Ashby plots for robots

• Increasing visibility of soft robotics as a field

  • Can we get grand challenges (i.e. Sandia fracture challenge) for soft robotics to help push the field?

  • Soft robotics curriculum

    • We need more multidisciplinary courses

    • The need for a soft robotics textbook: it is challenging because the field is rapidly changing

    • Undergraduate classes

  • Need to embed computational design for soft robotics into commercial CAD software to lower the entry bar so that everybody can use it effectively.

  • Accessibility of concepts