DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES
Ph.D. Thesis Colloquium
Speaker: Mr. Surya Ratna Prakash Dumpa
S.R. Number: 06-18-00-11-12-20-1-18701
Title: “Geometry-Consistent and Integrity-Aware Navigation for Autonomous Spacecraft under Partial Observability and Model Uncertainty”
Research Supervisor: Prof. Soumyendu Raha
Date & Time : May 28, 2026 (Thursday), 03:00 PM
Venue : #102, CDS Seminar Hall
ABSTRACT
Autonomous spacecraft navigation during powered descent and planetary landing operates under severe uncertainty, partial observability, and strict finite-time safety requirements. During such mission phases, navigation systems must reconcile nonlinear dynamics with noisy and incomplete measurements while remaining robust to structural model mismatch, estimator degradation, and transient instability. Classical filtering architectures, including Kalman filtering and standard particle filtering, primarily rely on posterior correction after state propagation and are typically analyzed using asymptotic stability notions. These approaches may become inadequate in autonomy-critical regimes where local geometric inconsistency, particle degeneracy, and short-horizon perturbation growth determine mission success before asymptotic behaviour becomes relevant.
This thesis develops a unified geometric and stochastic framework for integrity-aware autonomous spacecraft navigation in which geometric consistency, estimator integrity, probabilistic hazard forecasting, and finite-time transient stability are treated as interconnected components of the same problem.
First, a co-state-based differential–algebraic framework is developed to enforce local compatibility between system dynamics and measurement geometry under partial observability. The resulting geometric co-state variable acts as an intrinsic consistency signal linking nominal propagation with the measurement-consistent manifold. This signal is further used to construct stochastic regime abstractions and continuous-time hazard forecasting models, enabling probabilistic prediction of degraded operating modes through regime occupancy probabilities and mean first-passage indicators.
Second, an integrity-aware nonlinear filtering architecture, termed the Geometric Projection Particle Filter (GPF), is introduced for nonlinear and non-Gaussian state estimation under persistent model uncertainty. Unlike conventional bootstrap particle filters, the proposed method projects nominal drift dynamics directly onto the measurement-consistent subspace during propagation, creating a geometry-aware proposal process that reduces proposal–likelihood mismatch prior to weighting. This suppresses particle degeneracy, improves effective sample size retention, and maintains bounded estimation error in regimes where standard particle filters diverge.
Third, a logarithmic-norm-based framework is developed for analyzing finite-time transient stability of nonlinear It^o stochastic systems and projected data-constrained dynamics. By extending matrix measures in a Lipschitz sense, the approach provides tractable characterization of perturbation growth without requiring explicit Lyapunov function construction. The analysis clarifies the distinction between mean stability and pathwise finite-time safety, showing that stochastic diffusion induces nonzero probability of critical transient excursions even when average behaviour remains stable.
The proposed methods are validated using lunar lander descent navigation scenarios with partial observability and persistent model mismatch. Results demonstrate early hazard detection, improved estimator robustness, significant reduction in particle degeneracy, interpretable probabilistic hazard forecasting, and stronger finite-time safety characterization compared with classical filtering and asymptotic stability approaches.
These results establish that autonomous navigation should be treated not only as a state estimation problem, but as a coupled problem of geometric consistency, estimator integrity, stochastic hazard evolution, and finite-time operational safety. The thesis provides theoretical foundations and practical algorithmic realizations for integrity-aware autonomous spacecraft navigation.
ALL ARE WELCOME
