Robots are complex to design and produce, tricky to program, and have limited off-the-shelf configurations, with no option to physically adapt to their environments and specific tasks. In addition, the recent boom in additive manufacture presents roboticists with an incredibly high dimensional design space, but few tools that can effectively search the possibilities.
Evolution provides a bio-inspired pathway towards the autonomous design of robots that are environmentally specialized, and as such can be expected to outperform standard off-the-shelf solutions. Generational, iterative improvements, combined with environmental selection, allows robots to automatically develop their bodies and brains. In conjunction with the Autonomous Design Testbed in the Active Integrated Matter Future Science Platform, we are actively pursuing several evolutionary robotics themes:
- We can evolve robot end effectors, for example legs and arms, that are environmentally specialized and can be printed and attached to our robots on a per-mission basis for enhanced mission performance.
- Integration of evolutionary techniques to perform ‘design exploration’ in a space of possible robot configurations
- Evolutionary approaches Sim2real and reality gap
- Experimentation with soft robotic systems, which can morph and flex to navigate extreme environments
- Development of testbeds that allow for evolutionary robotics experimentation to occur reliably and repeatedly in hardware:
- a testbed for aerial robots, allowing mission-specific controllers to be evolved safely and repeatedly evolved for arbitrary UAVs with no human intervention required.
- A testbed that performs high-dimensional optimization on legged robots
- Automatic design SNN for control in hardware, e.g. on FPGAs
- Design of spiking ensemble controllers
CSIRO has developed a world-leading capability to evolve controllers on real robots, providing advanced mission outcomes for UAVs and hexapods. We are developing the capability to create task-specific robots for our clients. Our work has been featured on SCOPE, ABC, The Australian Financial Review, and in Wired magazine, with publications in top IEEE and ACM conferences, as well as Nature Machine Intelligence.
Paper: Comparing direct and indirect representations for environment-specific robot component design
Toward Singularity: a documentary about the raise of AI where Dr David Howard talks about Evolutionary Learning in robotics
Paper: Direct 3D Printing of Highly Anisotropic, Flexible, Constriction-Resistive Sensors for Multidirectional Proprioception in Soft Robots
Paper: Utilising Evolutionary Algorithms to Design Granular Materials for Industrial Applications