Visible to the public Learning for Control of Synthetic and Cyborg


We aim at enabling operation of synthetic and cyborg insects in complicated envi- ronments such as outdoors or inside collapsed buildings. Success in this research project will bring society closer to solving the grand challenge of having teams of mobile, disposable, search and rescue robots which can robustly locomote through uncertain and novel environments, finding survivors in disaster situations, while removing risk from rescuers. Learning and adaptation capabilities are critical to handle significant uncertainty in mobile platforms dynamics and environment vari- ables. As part of our cyborg insects thrust, we developed a flexible, multi-electrode polymer interfaces used to construct stable, chronic interfaces to the sensory organs of adult insects, and showed that the implanted interfaces can chronically record stimulus-triggered evoked potentials from eye and antennae afferents in both pupa and adult. Using a different interface, we ascertained that dragonflies show a turn response based on differential illumination of the three dragonfly ocelli. We are currently studying this response in tethered dragonflies with an aim to demonstrate turning behavior elicited in free flight. As part of our synthetic insects thrust, we instrumented an experiment are with an OptiTrack motion capture system to in- vestigate the gait properties of our new VelociRoACH crawler on three different surfaces: tile, gravel and carpet. Our goal is to identify efficient locomotive gaits for each terrain through reinforcement learning. Apart from dips at 4 and 7 Hz, we found that the robot's speed increases with stride frequency on all terrains. As part of our thrust towards collaborative exploration algorithms, we developed two new methods backed by strong theoretical guarantees. For the first, we challenge the expected return objective in risk-aware settings and we propose an new objec- tive inspired by Chernoff bounds. For the second, we show that adding a safety constraint designed to preserve ergodicity improves the efficiency of existing explo- ration algorithms. In both cases we complement the theoretical analysis with large scale experiments demonstrating computationally efficient planning algorithms.

Award ID: 0931463

Creative Commons 2.5

Other available formats:

Learning for Control of Synthetic and Cyborg