Safe Bipedal Locomotion and Navigation in Uncertain Environments
Loading...
Author(s)
Shamsah, Abdulaziz
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
This dissertation addresses the challenge of enabling bipedal robots to navigate and operate autonomously in dynamic and uncertain environments. While industrial robots have been highly successful in structured settings such as factories and warehouses---where robots perform predefined tasks in controlled environments with minimal uncertainty---replicating this success in real-world, dynamic scenarios remains a significant hurdle. Despite recent advancements in humanoid robotics, integrating these robots into unstructured environments is difficult due to the complexity of bipedal locomotion and the need to prioritize safety for both the robot and the surrounding agents.
This research focuses on overcoming these obstacles by developing a hierarchical approach that includes high-level task planning, mid-level motion planning, and low-level full-body control, with the overarching goal of ensuring safe navigation in uncertain environments. The two primary sources of uncertainty explored in this thesis are obstacle uncertainty, which involves dynamic obstacles such as autonomous grounded mobile robots and human pedestrians, and terrain uncertainty, which includes partial observability and unknown terrain elevation.
In this dissertation, we address the uncertainties in the following context. First, it addresses formal task and motion planning in partially observable environments with dynamic obstacles. In this work, the environment is partially observable when static obstacles occlude the robot’s view of certain regions in the environment; thus, guaranteeing collision avoidance with dynamic obstacles out of the robot's view becomes a challenging problem. To solve this, we develop a formal task planning framework based on Linear Temporal Logic (LTL), which provides guarantees for safe navigation and task completion. It employs a belief abstraction method to handle out-of-view dynamic obstacles. A key feature of this work is the abstraction of reduced-order model safety theorems into symbolic specifications to guarantee that the high-level task planner can be successfully executed by the underlying motion planner.
The second work addresses obstacle uncertainty in the context of social navigation. In this task, the bipedal robot is tasked to navigate and reach a specific goal in an open environment containing pedestrians. The bipedal robot is required to avoid collision with pedestrians and navigate in a socially acceptable manner. The pedestrians’ dynamics are not known, thus we introduce the Social Zonotope Network (SZN), a Conditional Variational Auto-encoder (CVAE) architecture for coupled pedestrian future trajectory prediction and ego-agent social path planning both parameterized as zonotopes. We integrate the SZN with a model predictive controller (MPC), where the zonotopes outputted by SZN are encoded as constraints for reachability-based motion planning and collision checking. Our results demonstrate the framework’s effectiveness in producing a socially acceptable path with consistent locomotion velocity and optimality.
Finally, we address terrain uncertainty in the context of search and rescue tasks, where we coordinate a heterogeneous team of bipedal and aerial robots. In this project, the terrain elevation is unknown. As the robots navigate the environment, they collect elevation data and update a terrain Gaussian process (GP) model. We present a terrain-aware MPC that solves the optimal paths for the bipedal while maximizing the traversability. We integrate lateral slopes derived from the terrain GP into the cost function of our proposed MPC framework. This method allows for a safer traversal of rough terrains by planning paths with minimum lateral slopes.
This dissertation contributes to advancing the field of bipedal robot navigation by providing a comprehensive set of tools and methods to address the challenges posed by dynamic and uncertain environments, offering solutions that ensure safety, optimality, and efficiency for real-world deployment.
Sponsor
Date
2024-12-04
Extent
Resource Type
Text
Resource Subtype
Dissertation