Date of Award
12-2025
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Automotive Engineering
Committee Chair/Advisor
Venkat Krovi
Committee Member
Phanindra Tallapragada
Committee Member
Matthias Schmid
Committee Member
Umesh Vaidya
Committee Member
Rahul Rai
Abstract
The skid-steered vehicle architecture has served as a useful vehicle platform for human-controlled operations. In particular, the absence of steering linkages and minimum moving parts makes this platform structurally-sound and scalable solution for operations in extreme environments. However, systematic modeling of the physics (kinematics/dynamics including the time-varying nonlinear skid-slip wheel terrain interactions) is crucial for realizing partial or complete automation. Thus, the primary motivation for this research is to enable autonomy development of skid-steered wheeled mobile robots (SSWMRs) with systematic investigation of the modeling/system-identification, creation of best-in-class simulation representations, and culminating in tradespace studies of both traditional and learning approaches for control. Additionally, we also seek to investigate applicability of AI/ML approaches and technologies that can systematically leverage multitudinous and copious streams of data to deduce patterns from the unstructured environments.
To this end, the overall problem has been investigated as two classes of problems: (a) finding environment-aware optimal navigation solutions; and (b) adapting the optimal solutions to changing environments. To further focus the efforts, a representative problem of skid-steered lane keeping in rough terrain is selected, without being overly limiting. Specifically, it highlights the need for awareness of human-imposed restrictions within the environment (e.g. lane bounds) while realizing optimal/adaptive solutions for varying environments and operational requirements.
In the first part, deep reinforcement learning (DRL) has been identified as a suitable candidate for identifying optimal vision-based control policies. Within this, the problem is centered much more around the systematic investigation of the tools that can be leveraged to minimize the simulation to reality (sim2real) transfer of the learnt policies. Primarily, these tools/methods can be segregated as techniques in problem formulation (such as state and action representations), reward-shaping and simulator calibration. The investigation is characterized by extensive experiments (both in simulation and on hardware) to validate the effectiveness of the proposed solutions. At the same time, it brings to sharp focus the necessity for large amounts of training data for learning generalizable policies that can be utilized in varying/alternate operation domains.
This limitation of the first part serves to motivate the second part of the efforts, that is, employing adaptive state estimation as a method to switch between varying operation environments. Numerous apriori methods to identify the operating conditions (off-line identification) and real-time operational-switching (online-adaption) have been investigated for SSWMRs. However, our efforts seek to develop a real-time online identification and adaptation framework for skid-steered operations. To this end, we adapt state-estimation methodologies to realize a light-weight and reliable solution to the problem of large-scale data requirement posed by the machine learning models. The investigation is supported by systematic experimental results and tradespace analyses for real-time operations of SSWMRs in changing environments and operational modes.
In combination, the two parts of this dissertation provision methods and tools for realization of a real-time optimal-adaptive framework for skid-steered vehicle operations in the context of the 4-wheeled SSWMR Husky robot. However, the developments of the second-part can be extended to a wider class of robotic-systems, which lack reliable analytical models but can leverage real-time sensing/measurement for data-driven modeling and control. This opens up possibilities for model-based control of broader class of robotic systems including aerial/marine vehicles, soft-robots and cable-driven parallel robots operating in challenging uncertain environments.
Recommended Citation
Salvi, Ameya, "Learning Enhanced System Identification and Control for Skid-Steered Robots" (2025). All Dissertations. 4160.
https://open.clemson.edu/all_dissertations/4160
Author ORCID Identifier
0000-0002-3307-5573