Algorithms and Architectures for Learning in-the-Loop Systems in Autonomous Flight

International Conference on Robotics and Automation (ICRA) 2019 Accepted Full-Day Workshop

Organizers

Dr. Aleksandra Faust, Senior Research Scientist, Google Brain, faust@google.com, http://www.afaust.info

Dr. Vijay Janapa Reddi, Associate Professor in Electrical Engineering, Harvard University, vjreddi@seas.harvard.edu, https://www.seas.harvard.edu/directory/vjanapa-reddi

Dr. Angela Schoellig, Assistant Professor, University of Toronto, schoellig@utias.utoronto.ca, http://www.dynsyslab.org/prof-angela-schoellig/

Dr. Sarah Tang, University of Pennsylvania/Nuro, Inc., sytang@alumni.seas.upenn.edu, http://www.sarahtang.net

Abstract

In past years, model-based techniques have successfully endowed aerial robots with impressive capabilities like high-speed navigation through unknown environments. However, task specifications, like goal positions, are often still hand-engineered.

To achieve true autonomy in applications such as disaster response and environmental-monitoring, unmanned aerial vehicles (UAVs) must additionally exhibit semantic understanding of tasks and environments, adaptation to unexpected changes, and robustness to unmodeled disturbances. Machine learning and deep learning have emerged as promising tools for higher-level autonomy, but are more difficult to analyze and implement in real-time.

Furthermore, maintaining high thrust-to-weight ratios for agility directly contradicts the need to carry sensor and computation resources, making hardware and software architecture equally crucial decisions.

This workshop aims to bring together researchers in the complementary fields of aerial robotics, learning, and systems to discuss the following themes:

• Learning for flight - How should learning be incorporated into UAVs’ perception-action loops?
• Structure in learning - How can models, structure, and priors enhance learning on UAVs?
• Performance guarantees - How can we analyze closed-loop performance of learning-in-the-loop systems
• Software+hardware co-design - How can we implement learning algorithms on resource-constrained UAVs? How should we simultaneously optimize algorithms and hardware choices to create lightweight, but highly-capable, UAVs?

Workshop Objectives

The objectives of this workshop are to:

• Discuss the aforementioned themes from the diverse perspectives of aerial robotics, learning, and systems researchers across academia, government, and industry
• Present state-of-the-art research results related to learning-in-the-loop for UAVs through keynote talks from leaders of related fields and contributed talks and posters
• Share details about, mistakes made, and “lessons learned” in hardware set-ups and software architectures that regular conference papers typically do not have space to expand on
• Stimulate and accelerate research in the learning for UAVs

We will host invited speakers that give a broad view of the state-of-the-art. This will include academic faculty as well as industry speakers on the commercial side and research side of UAV innovation.

Call for Papers

We are soliciting submissions of 4-page short papers (not including references) with up to a 2-minute accompanying video. Possible topics of interest include, but are not limited to:

• Combining model-based and model-free methods for autonomous flight
• Online learning and adaptation in mapping, perception, planning, and/or control for UAVs
• End-to-end learning of perception-action loops for flight
• Sample efficient learning on flying robots
• Learning for high-level autonomy in applications such as (but not limited to) disaster response, cinematography, search and rescue, environmental monitoring, aerial manipulation, agriculture, and inspection
• Closed-loop analysis learning-in-the-loop systems
• Metrics for evaluating the benefits of incorporating learning into perception-action loops or incorporating models into learning algorithms
• Challenges implementing learning algorithms in real-time on sensorimotor systems
• Novel architectures that use multi-agent networks or the cloud to decentralize demanding computations
• Insights into architecture design, system component choice, and implementation details (including “failed designs”) of real-time learning-in-the-loop algorithms

We strongly prefer work featuring experimental validation (including initial preliminary results) but will also consider simulation-only papers that convincingly address why the utilized simulator is a compelling representation of real-world conditions. We especially encourage papers that share valuable “failure analyses” or “lessons learned” that would benefit the community. We welcome work at all stages of research, including work-in-progress and recently accepted or published results.

Important Dates

Paper submission deadline: 24-Mar-2019

Author notification: 25-Apr-2019

Workshop: 23 or 24-May-2019