The objective of this research is to enable operation of synthetic and cyborg insects in complicated environments, such as outdoors or in a collapsed building. As the mobile platforms and environment have significant uncertainty, learning and adaptation capabilities are critical. The approach consists of three main thrusts to enable the desired learning and adaptation: (i) Development of algorithms to efficiently learn optimal control policies and dynamics models through sharing the learning and adaptation between various instantiations of platforms and environments. (ii) Creation of control learning algorithms which can be run on low-cost, low-power mobile platforms. (iii) Development of algorithms for online improvement of policy performance in a minimal number of real-world trials.
The proposed research will advance learning and adaptation capabilities of practical cyberphysical systems. The proposed approach will be generally applicable and lead to a new class of learning and adapting systems that are able to leverage shared properties between multiple tasks to significantly speed up learning and adaptation.
Success in this research project will bring society closer to solving the grand challenge of 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. This project will provide interdisciplinary training through research and classwork for undergraduate and graduate students in creating systems which intimately couple the cyber and physical aspects in robotic and living mobile platforms. Through the SUPERB summer program, under-represented students in engineering will experience research in learning and robotics.
This project targeted advancing the capabilities of synthetic and cyborg insects in complicated environments, such as outdoors or in a collapsed building. Research under this project consisted of three main thrusts: (i) Learning and adaptation for control to handle the significant uncertainty in these systems and their operating environments; (ii) Creation of synthetic, biomimetic insect-like robots; (iii) Investigation of bio-electrical interfaces in the context of cyborg insects. This project advanced the state of the art in learning and adaptation for control along two main directions: First, a new algorithm was developed to tackle the problem of planning in uncertain environments when risk aversion is desired --- i.e., when one doesn’t want to optimize the expected/average outcome, but rather optimize for the majority of the time the outcome being at least of a certain quality. Second, the problem of learning to control a system with (a priori) unknown dynamics was considered --- where the system is supposed to learn to control itself through experimentation. Two main challenges were studied, and addressed: How to perform this exploration safely, which is an important problem as exploration tends to pull the system towards the unknown, but moving into the unknown is often unsafe. How to extend existing exploration approaches that work well for (relatively small) discrete state space systems to continuous non-linear dynamical systems. This project advanced the state of the art in design methodology, rapid (and relatively cheap) fabrication techniques, and control of biomimetic milli-robots. We investigated the performance of a legged robot as it traversed three distinct rough terrains: tile, carpet, and gravel. Furthermore, we developed an accurate, robust, low-lag, and efficient algorithm for terrain classification that uses vibration data from the on-board inertial measurement unit and motor control data. We also developed a new ornithopter micro aerial vehicle, the 13 gram H2Bird. We augmented the ornithopter's built-in gyroscope-based control with a lightweight ground station, which was able to direct the ornithopter to a visually recognized goal, an open window. This project advanced the state of the art in understanding bio-electrical interfaces for cyborg insects. We have demonstrated a number of interface options enabling us to modulate insect flight. The first approach focused on identifying small, previously unstudied beetle muscles that, when stimulated would allow for fine-tuning of controlled turns. This approach also allowed us to highlight the benefit of the technology to fundamental studies of insect flight. In particular, the inability to record and exogenously stimulate neuromuscular signals during untethered flight has been a long-standing impediment to studies of natural insect flight. Using our system, we discovered a hidden function of a coleopteran flight muscle. A second direction focused on providing false information to insect sensory organs rather than simulating the "actuators"; we believe this approach will likely be more powerful and generalizable in the pursuit of efficient control algorithms. Most flying insects utilize two distinct visual subsystems during flight. The first is based on compound eyes. A second subsystem is comprised of three low resolution light sensors called ocelli. The ocelli are usually found on the insect face; differential changes in light flux falling on ocelli caused by attitude changes with respect to the horizon allow the insect to quickly compute pitch, roll and yaw. We have designed, built and tested ultra-miniature ocelli stimulators on dragonflies and locusts (Figure Y). The devices provide differential stimulation to ocelli in a package weighing less than 250 mg. Providing synthetic differentials in incident flux, either via absolute intensity control or pulse width modulation, between ocelli produces pronounced neck motions. This work has resulted in eleven publications. The PIs make their publications available to the public from their websites. Continued success along these lines of research will bring society closer to solving the grand challenge of 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. In addition to the pure research contributions, this project has provided several other contributions: This project has provided interdisciplinary training through research and classwork for twenty-one undergraduate students, eight graduate students, and two post-docs in creating systems which intimately couple the cyber and physical aspects in robotic and living mobile platforms. Each year the PIs have held open-labs, hosting hundreds of visitors (primarily parents with K-12 kids and prospective freshmen)---aiming to attract the next generation to STEM disciplines. The PIs have also combined for over a dozen public lectures about the work under this project. Maharbiz has begun to write about the ethical implications of developing autonomous systems based on living insects (and similar organisms). Along with his group, he has started a site devoted to discussing these and related issues (http://the-interface.org/).
