Data-driven predictive pursuit-evasion engagement guidance and fast posture reconstruction of soft continuum arm

Akcal, Ugur

Permalink

https://hdl.handle.net/2142/129212 Copy

Description

Title

Data-driven predictive pursuit-evasion engagement guidance and fast posture reconstruction of soft continuum arm

Author(s)

Akcal, Ugur

Issue Date

2025-04-17

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

Chowdhary, Girish

Department of Study

Siebel School Comp & Data Sci

Discipline

Computer Science

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Artificial Neural Networks
• Data-driven Predictive Guidance
• Posture Reconstruction
• Soft Continuum Arm

Language

eng

Abstract

This thesis explores artificial neural network-based methodologies designed to improve performance in two critical areas of robotics: high-precision pursuit-evasion guidance and soft continuum arm posture reconstruction. First, a predictive guidance scheme is developed to enable rapid interception of agile and evasive targets with limited knowledge of evader dynamics. A recurrent neural network is trained on representative evader maneuvers to efficiently predict future acceleration commands. These predictions are incorporated into a finite-horizon optimal control problem, generating near-optimal guidance commands that significantly outperform traditional reactive laws such as proportional navigation in dynamic engagement scenarios, particularly in terms of average miss distance. Second, the thesis introduces a framework in which the Vicon motion capture system is leveraged to acquire high-fidelity ground-truth posture data in order to train an artificial neural network for fast and smooth posture reconstruction of soft continuum arms. Given the infinite-dimensional nature of soft-arm deformation, strain fields are represented using a low-dimensional set of principal components. A feed-forward neural network is trained in an unsupervised manner with a physics-informed loss to instantly infer the coefficients for the principal components from sparse marker measurements. This approach allows for real-time posture reconstruction, achieving computation speeds five orders of magnitude faster than classical iterative or optimization-based techniques while preserving accuracy and smoothness. Together, these contributions underscore the potential of neural networks to unify control, estimation, and efficient computation in robotics. By bridging pursuit-evasion engagements and continuum robot shape reconstruction, the thesis highlights the versatility and performance gains afforded by data-driven models, ultimately paving the way for advanced, high performance robotic autonomy in both aerial and soft arm applications.

Graduation Semester

2025-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129212

Copyright and License Information

Copyright 2025 Ugur Akcal

Owning Collections