Efficient algorithms for risk-averse air-ground rendezvous missions

Barsi Haberfeld, Gabriel

Permalink

https://hdl.handle.net/2142/113913 Copy

Description

Title

Efficient algorithms for risk-averse air-ground rendezvous missions

Author(s)

Barsi Haberfeld, Gabriel

Issue Date

2021-12-03

Director of Research (if dissertation) or Advisor (if thesis)

Hovakimyan, Naira

Doctoral Committee Chair(s)

Hovakimyan, Naira

Committee Member(s)

• Salapaka, Srinivasa
• Stipanovic, Dusan
• Voulgaris, Petros
• Theodorou, Evangelos

Department of Study

Mechanical Sci & Engineering

Discipline

Mechanical Engineering

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Planning
• risk
• optimization
• model predictive control
• optimal control
• logistics

Abstract

Demand for fast and inexpensive parcel deliveries in urban environments has risen considerably in recent years. A framework is envisioned to enforce efficient last-mile delivery in urban environments by leveraging a network of ride-sharing vehicles, where Unmanned Aerial Systems (UASs) drop packages on said vehicles, which then cover the majority of the distance before final aerial delivery. By combining existing networks we show that the range and efficiency of UAS-based delivery logistics are greatly increased. This approach presents many engineering challenges, including the safe rendezvous of both agents: the UAS and the human-operated ground vehicle. This dissertation presents tools that guarantee risk-optimal rendezvous between the two vehicles. We present mechanical and algorithmic tools that achieve this goal. Mechanically, we develop a novel aerial manipulator and controller that improves in-flight stability during the pickup and drop-off of packages. At a higher level and the core of this dissertation, we present planning algorithms that mitigate risks associated with human behavior at the longest time scales. First, we discuss the downfalls of traditional approaches. In aerial manipulation, we show that popular anthropomorphic designs are unsuitable for flying platforms, which we tackle with a combination of lightweight design of a delta-type parallel manipulator, and L1 adaptive control with feedforward. In planning algorithms, we present evidence of erratic driver behavior that can lead to catastrophic failures. Such a failure occurs when the UAS depletes its resource (battery, fuel) and has to crash land on an unplanned location. This is particularly dangerous in urban environments where population density is high, and the probability of harming a person or property in the event of a failure is unsafe. Studies have shown that two types of erratic behavior are common: speed variation and route choice. Speed variation refers to a common disregard for speed limits combined with different levels of comfort per driver. Route choice is conscious, unconscious, or purely random action of deviating from a prescribed route. Route choice uncertainty is high dimensional and complex both in space and time. Dealing with these types of uncertainty is important to many fields, namely traffic flow modeling. The critical difference to our interpretation is that we frame them in a motion planning framework. As such, we assume each driver has an unknown stochastic model for their behavior, a model that we aim to approximate through different methods. We aim to guarantee safety by quantifying motion planning risks associated with erratic human behavior. Only missions that plan on using all of the UAS's resources have inherent risk. We postulate that if we have a high assurance of success, any mission can be made to use more resources and be more efficient for the network by completing its objective faster. Risk management is addressed at three different scales. First, we focus on speed variation. We approach this problem with a combination of risk-averse Model Predictive Control (MPC) and Gaussian Processes. We use risk as a measure of the probability of success, centered around estimated future driver position. Several risk measures are discussed and CVaR is chosen as a robust measure for this problem. Second we address local route choice. This is route uncertainty for a single driver in some region of space. The primary challenge is the loss of gradient for the MPC controller. We extend the previous approach with a cross-entropy stochastic optimization algorithm that separates gradient-based from gradient-free optimization problems within the planner. We show that this approach is effective through a variety of numerical simulations. Lastly, we study a city-wide problem of estimating risk among several available drivers. We use real-world data combined with synthetic experiments and Deep Neural Networks (DNN) to produce an accurate estimator. The main challenges in this approach are threefold: DNN architecture, driver model, and data processing. We found that this learning problem suffers from vanishing gradients and numerous local minima, which we address with modern self-normalization techniques and mean-adjusted CVaR. We show the model's effectiveness in four scenarios of increasing complexity and propose ways of addressing its shortcomings.

Graduation Semester

2021-12

Type of Resource

Thesis

Permalink

http://hdl.handle.net/2142/113913

Copyright and License Information

© 2021 Gabriel Barsi Haberfeld

Owning Collections