ST3M Method Overview

• ST3M method is a dual-application framework that unifies reinforcement learning for VTOL UAV trajectory tracking and PCA-GP-based STEM distortion correction.
• For UAV applications, it leverages RL and optimized path planning to achieve smoother transitions with lower attitude excursions and reduced tracking errors.
• In STEM, the method applies PCA to isolate drift and uses Gaussian process regression to correct spatial and temporal distortions with sub-pixel precision.

The acronym ST3M refers to distinct domain-specific methodologies in recent literature. Here, the primary contexts are: (1) Smooth Transition Trajectory Tracking Method for reinforcement learning-based control of VTOL UAV transitions, and (2) a Gaussian-process regression-based pipeline for spatial and temporal distortion correction in scanning transmission electron microscopy (STEM). Each is detailed below with technical rigor appropriate for advanced research professionals.

1. Mathematical Formulation and Problem Contexts

1.1. ST3M for VTOL UAVs

The ST3M framework for VTOL UAVs addresses transition control—specifically, the trajectory tracking of vehicles as they move between hover and cruise regimes. The problem is cast as a reference-trajectory tracking task, formalized as minimizing a quadratic cost function over a sequence of “hover points,” with each point given by $(P_d^i, I_d^i)$. The optimization per trajectory segment enforces smoothness via constraints on velocity and angular increments: $\min_{P^i, I^i} \;\; J = \sum_{i=1}^N e^i$

$$e^i = \begin{bmatrix} P^i - P_d^i\ I^i - I_d^i \end{bmatrix}^T \begin{bmatrix} Q_1 & 0\ 0 & Q_2 \end{bmatrix} \begin{bmatrix} P^i - P_d^i\ I^i - I_d^i \end{bmatrix}$$

subject to: $$0 \leq \|\dot{P}^{i+1} - \dot{P}^i\| \leq \epsilon_1^{\max}, \quad 0 \leq \|\dot{I}^{i+1} - \dot{I}^i\| \leq \epsilon_2^{\max}$$ where $P$ is position, $I$ is the vector of Euler angles, $Q_1$, $Q_2$ are weight matrices, and $\epsilon^{\max}$ bounds the increment rates.

1.2. ST³M in STEM

ST³M in the electron microscopy domain denotes a method for learning and correcting spatial and temporal distortions in STEM via multivariate analysis of atomic trajectories and Gaussian process regression. The distortion mapping is defined as: $$D(x', y', t) = [x(x', y', t) - x', \;\; y(x', y', t) - y'] = (\Delta x, \Delta y)$$ with a decomposition: $$D(x', y', k) = D_{\rm stat}(x', y') + \delta D(x', y', k)$$ which separates time-stationary effects from slowly varying temporal distortions.

2. Core Algorithmic Steps and Theoretical Frameworks

2.1. UAV ST3M: Cruise-as-Hover Paradigm

The ST3M algorithm interprets cruise flight as “dynamic hover” at variable tilt angles ($\mu$). Rather than discrete mode switching, a unified RL policy is trained to stabilize and move the vehicle across the entire tilt space ($\mu\in[0,90^\circ]$). The core steps are:

• Reinforcement learning of a hover-mode RL agent (HRM).
• Path planning (BalancePath, improved A*) to generate points for the UAV to “try to hover” along, forming a smooth transitional sequence.

2.2. STEM ST³M: PCA-GP Distortion Correction

The ST³M workflow involves:

• Extraction of atomic column trajectories frame-wise from a STEM stack using a U-net neural network.
• Principal component analysis (PCA) of trajectories to isolate rigid drift (first PC) and higher-order distortions.
• Gaussian process regression (GPR) to interpolate the residual distortion field per coordinate, optimizing hyperparameters via log marginal likelihood.

3. Reinforcement Learning and Statistical Learning Details

Approach	State/Input Space	Action/Output Variables	Learning Mechanism	Loss/Objective
UAV ST3M	$$O = [I, 0.25\,\sigma(V),\sigma(\omega), \Delta P, S_{\rm prev}]$$	$[\Omega_1, \Omega_2, \Omega_3, \mu_1, \mu_2]$	PPO (Proximal Policy Optimization)	Policy gradient (PPO)
STEM ST³M	Atomic column trajectories (PCA-filtered)	Distortion field $D(x', y', t)$	GP regression with RBF kernel	Log marginal likelihood

In the UAV context, the reward signal at each step aligns the body z-axis with earth-z, penalizes velocities, and is governed by tunable weights $W_V, W_\omega$. Progressive curriculum training is used to improve hover-to-cruise tracking via increasing target displacement magnitudes. In STEM, PCA disentangles global drift from local/nonlinear artifacts before these are addressed by GP.

4. Architectural and Implementation Aspects

4.1. UAV Coupled Control

The RL agent in ST3M outputs both rotor speeds and tilt angles, fusing thrust magnitude (altitude) and vectoring (horizontal translation) in a single policy. This enables simultaneous control over position and attitude without manual PID tuning.

4.2. STEM Data Analysis Pipeline

The computational workflow includes:

• Data ingestion (e.g., using pydm3reader), atom localization (UNet in PyTorch), Gaussian-peak refinement, and PCA (Scikit-learn).
• GP regression (Pyro or GPyTorch).
• Backward warping of images using per-pixel correction maps (map_coordinates in SciPy).

Implementation Snippets

import pyro.contrib.gp as gp kernel = gp.kernels.RBF(input_dim=3, variance=1.0, lengthscale=torch.tensor([10,10,5])) gpr = gp.models.GPRegression(X_train, y_train, kernel, noise=torch.tensor(0.01)) gpr.optimize() mu, cov = gpr(X_grid, full_cov=True)

5. Empirical Results and Comparative Metrics

UAV Trajectory Tracking

ST3M demonstrates lower attitude excursions and position errors compared to classical dual-loop PID controllers in simulation:

• Mean pitch: PID $=47.57^\circ$, ST3M $=20.25^\circ$.
• Max position error: PID $=4.726$ m, ST3M $=2.648$ m.
• Mean position error: PID $=2.357$ m, ST3M $=0.861$ m.

Real-world experiments confirm smooth cruise-to-hover transitions with seamless attitude profiles and no explicit mode transitions (Lin et al., 3 Dec 2025).

STEM Distortion Correction

ST³M achieves sub-pixel calibration and correction of STEM images, with PCA enabling effective separation of drift and higher-order distortions, and GP quantifying spatial/temporal uncertainty and smoothness (Roccapriore et al., 2020).

6. Limitations, Extensions, and Theoretical Considerations

Limitations

• UAV ST3M requires a high-fidelity digital twin and careful reward shaping; scaling in curriculum complexity can affect training time; model-free RL lacks formal stability guarantees of MPC.
• STEM ST³M faces $\mathcal{O}(M^3)$ cubic GP complexity; standard kernels (RBF) assume smoothness, which may not suit all experimental artifacts. Homoscedastic noise is assumed, potentially requiring adaptation for per-column uncertainty.

Potential Remedies and Extensions

• UAV ST3M can be extended to wind/disturbance robustness (domain randomization), more complex waypoint/path constraints, safe-RL/Lyapunov-inspired critics for partial guarantees, and multi-agent or fault-tolerant scenarios.
• STEM ST³M can use sparse GPs, structured kernels, or mini-batch SVI to address computational scaling; alternative kernels can improve handling of discontinuous artifacts.

7. Significance and Applications

ST3M methodologies provide unified, learning-based frameworks that replace rigid mode-switching and decoupled control paradigms with continuous, holistic, and data-driven control laws (UAV) or correction pipelines (STEM). In UAVs, this results in reduced vibration and improved tracking through a single controller. In electron microscopy, ST³M supports robust, in-line spatial–temporal calibration for quantitative atomic imaging, enabling sub-pixel uncertainty quantification and flexible correction of both stationary and non-stationary distortions without strong modeling assumptions.

References:

• "A Learning-based Control Methodology for Transitioning VTOL UAVs" (Lin et al., 3 Dec 2025)
• "Identification and Correction of Temporal and Spatial Distortions in Scanning Transmission Electron Microscopy" (Roccapriore et al., 2020)