Is this project an undergraduate, graduate, or faculty project?
Graduate
Project Type
individual
Campus
Daytona Beach
Authors' Class Standing
Emre Girgin, Graduate student
Lead Presenter's Name
Emre Girgin
Lead Presenter's College
DB College of Engineering
Faculty Mentor Name
Dr. Cagri Kilic
Abstract
Autonomous exploration of extraterrestrial environments using legged robots requires robust GNSS-free 3D state estimation. Standard leg odometry relies on Zero-Velocity Updates (ZUPT), which assume a grounded foot remains completely stationary. This assumption consistently fails on deformable granular terrain due to unobservable slippage, rapidly degrading state estimation. To mitigate this critical failure mode, we propose a dual contact-detection framework designed to robustly gate an Error-State Extended Kalman Filter (ESEKF) tracking pose, velocity, and IMU biases. The architecture isolates physical load and kinematics by modeling contact detection as two independent parallel Hidden Markov Models (HMMs). The Load HMM processes Ground Reaction Forces, while the Kinematics HMM processes estimated foot velocity and position derived from forward kinematics. Emission probabilities for both models are evaluated using a bimodal Gaussian Mixture Model to differentiate between the swing phase, characterized by zero-mean and low-variance, and the stance phase, characterized by high-mean and high-variance distributions. To handle dynamic terrain interactions, the Kinematics HMM replaces traditional static transition probabilities with a dynamic, energy-based matrix. State-switch likelihoods are scaled exponentially by foot kinetic energy and adapted online via cross-entropy loss against a dual-HMM fused label. At each time step, modality beliefs are updated via a recursive Forward Filter. The independent outputs are fused using a logical AND gate for the discrete state estimation and a normalized probabilistic product for the final continuous stance belief. Testing validates this framework over hundreds of meters against baseline methods relying on simple load thresholding. On granular media, the approach reduces Absolute Trajectory Error (ATE) by up to 70% and trajectory drift by 84%. On deformable surfaces, it achieves 74% and 83% reductions, respectively. This energy-modulated, dual-modal architecture effectively mitigates unobservable terrain-yielding drift, ensuring rigorous state estimation stability necessary for autonomous extraterrestrial navigation.
Did this research project receive funding support (Spark, SURF, Research Abroad, Student Internal Grants, Collaborative, Climbing, or Ignite Grants) from the Office of Undergraduate Research?
No
Included in
Energy-aware Bimodal Contact Detection for Leg Odometry
Autonomous exploration of extraterrestrial environments using legged robots requires robust GNSS-free 3D state estimation. Standard leg odometry relies on Zero-Velocity Updates (ZUPT), which assume a grounded foot remains completely stationary. This assumption consistently fails on deformable granular terrain due to unobservable slippage, rapidly degrading state estimation. To mitigate this critical failure mode, we propose a dual contact-detection framework designed to robustly gate an Error-State Extended Kalman Filter (ESEKF) tracking pose, velocity, and IMU biases. The architecture isolates physical load and kinematics by modeling contact detection as two independent parallel Hidden Markov Models (HMMs). The Load HMM processes Ground Reaction Forces, while the Kinematics HMM processes estimated foot velocity and position derived from forward kinematics. Emission probabilities for both models are evaluated using a bimodal Gaussian Mixture Model to differentiate between the swing phase, characterized by zero-mean and low-variance, and the stance phase, characterized by high-mean and high-variance distributions. To handle dynamic terrain interactions, the Kinematics HMM replaces traditional static transition probabilities with a dynamic, energy-based matrix. State-switch likelihoods are scaled exponentially by foot kinetic energy and adapted online via cross-entropy loss against a dual-HMM fused label. At each time step, modality beliefs are updated via a recursive Forward Filter. The independent outputs are fused using a logical AND gate for the discrete state estimation and a normalized probabilistic product for the final continuous stance belief. Testing validates this framework over hundreds of meters against baseline methods relying on simple load thresholding. On granular media, the approach reduces Absolute Trajectory Error (ATE) by up to 70% and trajectory drift by 84%. On deformable surfaces, it achieves 74% and 83% reductions, respectively. This energy-modulated, dual-modal architecture effectively mitigates unobservable terrain-yielding drift, ensuring rigorous state estimation stability necessary for autonomous extraterrestrial navigation.