Research

Causal Structure Recovery Under Partial Observability: Why Filtered States Outperform RTS-Smoothed States for Granger Discovery

ORIA Research

The standard estimate-then-discover pipeline is systematically biased when built on RTS-smoothed states; filtered state estimates preserve the exogeneity conditions for Granger-style discovery and recover near-oracle structure at no extra cost. Accepted to ICAI’26 (Springer Nature); full paper to be linked upon publication.

• Causal inference
• State-space
• Time series
• Granger
• Kalman

Causal-Reliability-Aware State Estimation: Learning Admissible Representations under Partial Observability

ORIA Research

The constructive sequel to our filtering-vs-smoothing diagnostic: rather than asking which estimator is better, it asks which estimated-state representations are admissible for causal discovery at all. The answer rests on two independent requirements, each backed by a verified counterexample. Self-consistency: a proxy must follow its own white-residual recursion, or moving-average structure aliases into spurious autoregressive edges. Coordinate alignment: even a perfectly whitened, exogenous proxy recovers a basis-scrambled graph unless its coordinates are pinned to physical sensors — the classical rotation indeterminacy. The proposed architecture, SCEM, enforces both through a structural firewall separating a strictly forward, graph-readable causal channel from a bidirectional channel dedicated to reconstruction. In well-specified linear-Gaussian systems the causal channel provably reduces to the Kalman filter, so SCEM preserves filter-grade causal accuracy while matching smoother-grade reconstruction error — resolving the accuracy-vs-admissibility trade-off by decomposition instead of compromise. Accepted to IEEE MLSP 2026; full paper to be linked upon publication.

• Causal discovery
• Kalman filtering
• Representation learning
• Identifiability
• Partial observability

Posture-Aware Active Shelf Inspection under Velocity–Posture Tracking Constraints

ORIA Research

Capstone of ORIA’s Onboard volume on legged active perception. On a quadruped, body height and pitch reshape the sensor viewpoint, so the planner selects velocity, height, and pitch to reduce shelf-occupancy uncertainty while staying inside the velocity–posture command region the controller can actually track — the achievable-viewpoint envelope. Sweeping 270 commands for a simulated Unitree Go2, only 86 are trackable, concentrated at moderate heights and failing at extreme crouch and at high-height, high-speed commands. Under matched viewpoints, non-nominal posture exposes 13–17.5 percentage points more audit cells within the reach window, and across five layouts a posture-enabled planner removes 42% of change-risk-weighted entropy versus 26% at nominal posture. A myopic (H=1) planner performs comparably, indicating that posture availability — not planning horizon — accounts for most of the gain. Accepted to ICAI’26 (Springer Nature); full paper to be linked upon publication.

• Active perception
• Legged robots
• Viewpoint planning
• Posture control
• Information gain
• Warehouse inspection

Modeling the Impact of Partial Shading on Photovoltaic Panels

Partial shading is a critical challenge for PV systems, causing power loss and damaging hotspots; existing detection relies on expensive hardware or oversimplified analytical models that fail under dynamic conditions. This work presents an end-to-end machine learning approach for predicting shading percentage via regression. A custom MATLAB Simulink physics-based single-diode equivalent-circuit model generated >24 million data points, varying irradiance (200–1000 W/m²), temperature (10–50 °C), twelve series-parallel array configurations, and shading levels (0–100%). A context-aware feature-engineering step normalizes voltage, current, and power into scale-invariant ratios so one model generalizes across configurations. The model uses XGBoost regression to estimate PV output under partial shading, demonstrating that properly processed I–V curve data alone can quantify shading without costly specialized hardware — a practical tool for optimizing residential PV systems.

• Photovoltaics
• Partial shading
• XGBoost
• Regression
• Simulation dataset

Innovative Applications of AI and 3D Printing in Digital Dentistry: Enhancing Accuracy and Efficiency in Dental Care

Digital dentistry integrates digital technologies (CAD/CAM, 3D printing, AI) into diagnosis, treatment planning, and restoration. 3D printing in dentistry still has drawbacks: complex design, reliance on manual adjustments, limited resolution with occasional inaccuracies. The authors propose an AI-enabled framework: import an STL file, select the application, and begin printing within seconds; the system analyzes the 3D object, adjusts orientation, and generates supports for splints, crowns, bridges, and surgical guides. AI enhances precision, efficiency, and reliability by automating previously manual, error-prone steps — detecting complex geometries, optimizing support structures, and adjusting printing parameters in real time — reducing design iterations while improving resolution and accuracy.

• 3D printing
• Digital dentistry
• Autoencoder
• Support optimization
• CAD/CAM

Study on the Use of Random Forest Classifier and Multi-Output Classifier Models for Predicting Student Academic Performance and Identifying Areas of Concern

Explores using machine learning to identify factors connected to student academic performance and to predict learning outcomes early, using two models: Random Forest classifier and Multi-Output classifier. Random Forest builds multiple decision trees and takes the mode of predictions, identifying the most significant factors. The Multi-Output classifier (neural-network backend) handles multi-label tasks, predicting several targets at once (e.g., grade and engagement level). Using several Kaggle student datasets, Random Forest identified key factors (units completed prior semester, prior-semester grades, tuition fee payment status) with 85.9% accuracy (94.5% correct on non-dropouts, 67.9% on dropouts). The Multi-Output Classifier yielded 83.5%–94.2% accuracy across five target variables — giving educators insight to advocate tailored support for at-risk students.

• Random forest
• Neural networks
• Multi-output classifier
• Early-warning systems
• Education

Local Classification is Insufficient: Recovering Lecture Structure Through Temporal Constraint Enforcement

ORIA Research

The research strand of the Chalkless initiative — recovering the logical flow of a lecture (how definitions, examples, and key concepts build over time) beyond the words spoken; combines multimodal capture with sequence modeling to reconstruct a lecture’s reasoning arc in real time.

• Lecture analysis
• Multimodal
• Sequence structure
• Temporal rules
• Educational AI

Gait-Conditioned Multi-View LiDAR Reconstruction for Occluded Shelf Inspection

ORIA Research

Paper I of ORIA’s Onboard volume on legged active perception for warehouse inventory (Unitree Go2). Rather than treating a quadruped’s gait-induced bobbing, pitch, and roll as noise to be filtered, this work exploits locomotion as a controlled source of viewpoint diversity: motion-corrected LiDAR sweeps are aggregated across gait phases to build a probabilistic visibility field that recovers deep-shelf geometry occluded from any single static scan. The central claim is observability — which shelf volumes become visible because of motion — rather than deblurring, validated in Isaac Lab against static and stabilized-trajectory baselines. Full paper linked once it completes review.

• Quadruped
• LiDAR
• Active perception
• Visibility field
• Occlusion recovery

Volatility-Aware Semantic Pose-Graph SLAM for Repetitive, Dynamic Warehouse Aisles

ORIA Research

Paper II of the Onboard volume. Warehouses break classical SLAM through geometric repetition (identical aisles alias into false loop closures) and volatility (inventory churns hourly). This work models landmark permanence as an inferred latent state that evolves over time and space — a spatio-temporal permanence field — so the factor graph trusts structural anchors over transient stock, and validates loop closures against bay-level semantic-topological signatures rather than single-landmark geometry. Weighting becomes a consequence of estimated structural reliability, not a fixed semantic lookup. Full paper linked once it completes review.

• SLAM
• Landmark permanence
• Loop closure
• Pose graph
• Perceptual aliasing

Project CLOSR: Causal Lunar Operations & Sovereign Reasoning Engine

ORIA Research

An edge-native causal world model for autonomous planetary robotics, aimed at AMD’s Advancing AI 2026 Physical AI & Robotics track. CLOSR ingests high-resolution NASA LRO/LOLA lunar topography together with rover telemetry into a Spatio-Temporal Causal State Space Model that structurally separates passive environmental transitions from active robotic interventions. An on-board counterfactual query engine evaluates Pearl’s do-operator in real time, letting the agent isolate internal mechanical faults (e.g., motor-winding failure) from external constraints (e.g., regolith wheel slip) before actuating and propose safer alternative trajectories — all authored in PyTorch and optimized for low-latency local inference on AMD ROCm hardware. Full paper linked once it completes review.

• Causal world model
• Edge AI
• Space robotics
• Counterfactual reasoning
• AMD ROCm