Logan Luna

Research

Three threads of work, with the figures, papers and code behind each.

AI for Mathematical Reasoning

My central question is how semantically informative training signals can be quantified and incorporated into the training and generation of mathematical reasoning models. Symbolic feedback rewards output that compiles or type-checks, which does not certify that it expresses the intended mathematics. Current directions:

  1. Extending reinforcement learning from symbolic feedback with semantic alignment as the training objective, measured by pass@k at large k.
  2. Dense rewards from distance in a learned proof-semantic space, grounded in kernel-checked entailment.
  3. Diffusion language models as a generative substrate for theorem proving and auto-formalization.

AIProver: Agentic Auto-Formalization of Mathematical Research via Certificate-Driven Evolving Harness

Under reviewICLR 2027

Prithwish Jana, Viet Bach Hoang, Logan Luna, Viresh Pati, Akash Singirikonda, Cy Xie, Lisa Carbone, Wuyang Chen, Walter Moreira, Joe Stubbs, Sriram Vishwanath, Vijay Ganesh.

Contribution: RL pipeline and reward, BEq+ extension, distillation pipeline and dataset, Semantic Alignment Model. Results: pass@4 semantic correctness from 15.7% to 36.7% standalone (above GPT-5.6-Sol and Codex); 41.9% to 79.8% as a Claude Code skill.

Fig. 1. An illustrative AIProver session: an English theorem becomes Lean, is repaired from Lean's own error with a Mathlib lemma, and is checked against the English. Above the file, the training stage behind each step lights up; choose one to see it.

Orchestration run replay

A real run of the orchestration, replayed from its trace: formalization, a 20-lemma sketch, AIProver jobs checked by Lean, and the lemmas that were handed back to the orchestrator. Conditional on the smooth shifted primes hypothesis (H).

Erdős's conjecture on large totient fibers

Full replay →
Fig. 2. A real 21-hour orchestration run, replayed from its trace on a compressed clock: each dot is a lemma, filling green as Lean accepts AIProver's proof.

ATLAS: Autoformalized Textbook Library At Scale

Meta FAIR, Research Collaborator V

A verified Lean 4 library built with AutoformBot, an open-source multi-agent framework that translates informal textbook mathematics into machine-checked definitions and proofs. I contribute to ATLAS in a contract role supporting Meta FAIR (Core Learning and Reasoning).

Reinforcement Learning for Safety-Critical Systems

Earlier work in reinforcement learning for safety-critical systems: satellite collision avoidance, intrusion detection, and sim-to-real drone control.

Satellite Trajectory Optimization via Proximal Policy Optimization for Space Debris Avoidance

JournalIEEE Access, 2026

Logan Luna, Juan Ortiz Couder, Raul Alejandro Vargas-Acosta.

97.5% collision-avoidance success over 1,000 deterministic GEO episodes, 76.8 and 70.0 percentage points above the rule-based and impulsive Δv baselines.

Globe surrounded by orbiting debris fragments from the Fengyun-1C, Cosmos 2251 and Iridium 33 breakups
Fig. 3. Tracked debris from the Fengyun-1C, Cosmos 2251 and Iridium 33 breakups, with an illustrative avoidance manoeuvre magnified in the inset.

AERO-Sim2Real: Reinforcement Learning for Sim-to-Real Drone Deployment

PosterSenior Design, Embry-Riddle, 2026

Logan Luna, Sirio Jansen-Sánchez, Vincent Reggi, Kolton Galie, Nicolas Andricos.

PPO policies trained in NVIDIA Isaac Lab and flown autonomously on Crazyflie 2.1 hardware; hover transferred at 1.3/2.1/0.7 cm real-vs-sim RMSE, while a figure-eight task isolated aggressive horizontal maneuvering as the dominant transfer gap.

Dueling Deep Q-Learning for Intrusion Detection

ConferenceIEEE SoutheastCon 2025, Best Paper Award

Logan Luna, Matthew P. Berkowitz, Laxima Niure Kandel, Sirio Jansen-Sánchez.

99.68% multi-class accuracy on 2.18M CIC-IDS2018 flows, against 88% for prior DQN-based detection and 98.4–99% for tree-ensemble and random-forest IDSs; SHAP attribution of the learned policy.

Fig. 4. Flows pass an XGBoost screen; malicious ones run through the dueling Q-network, whose value and advantage rejoin as one Q-value per attack class. Recall is the real CIC-IDS2018 result; the flows are illustrative.

Direct Policy Transfer from Simulator to Firmware: Verification Boundaries for Learned UAV Autonomy

Under reviewACM International Conference on Trustworthy and Responsible AI and Computing Systems

Illia Biblyi, Elsy Fuentes Montiel, Andressa C. Ojeda, Logan Luna, Kolton Galie, Vincent Reggi, Nicolas Andricos, Sirio Jansen-Sánchez, M. Ilhan Akbas.

A verification framework for compiling Isaac Lab policies directly into Crazyflie nano-UAV firmware, with executable checks from observation replay to source-to-binary traceability; 2–3 cm hover altitude error onboard.

Fig. 5. One policy flies the same plan in simulation and on the Crazyflie. The real drone's drift is drawn four times larger; the inset gives it in centimetres. Illustrative tracks.

Sensing and Scientific Workflows

Earlier work in signal-centric remote sensing and FAIR scientific workflows.

Signal-Centric Remote Sensing via Alternative Preprocessing and Acoustic Processing for ML-Driven Applications

ConferenceIEEE SoutheastCon 2025, Concord, NC

Logan Luna, Sirio Jansen-Sánchez, Ilteris Demirkiran, Leo Ghelarducci.

91.2% reduction in total sonar processing time (13.1% in filtering alone) against image-based CV processing, while preserving target structure the image pipeline removed.

Fig. 6. One sonar sweep past an underwater gate, cleaned two ways. Filtering the raw signal keeps the gate echo in 3.1 ms; rendering images first takes 37.4 ms and erases it. Waterfalls and timings from the paper, slowed down.

SensorWF: A FAIR Generalizable Workflow Framework for Scientific Time-Series Analysis

WorkshopIEEE eScience 2026 (ReWorDS Workshop), Naples, Italy

Logan Luna, Joseph Rigo, Kellan Shew, Raul Alejandro Vargas-Acosta.

One analytical core carried across spacecraft telemetry, ECG, and climate data with small domain adapters and no core changes; bit-identical outputs under fixed seeds.

Fig. 7. Real spacecraft, ECG and weather recordings, each read by its own small adapter (M1), then marked by the same unchanged modules M2 to M5.

Code

  • Milo: Magnitude-Invariant Learning Optimizer

    A group-wise gradient normalization optimizer designed to reduce magnitude distortion in learning dynamics, tested across image classification, CNNs, and reinforcement learning tasks.

  • SensorWF: A FAIR Generalizable Workflow Framework for Scientific Time-Series Analysis

    A FAIR workflow framework that carries one analytical core across spacecraft telemetry, ECG, and climate data via small domain adapters, with bit-identical outputs under fixed seeds (IEEE eScience 2026, ReWorDS Workshop).

  • Learning Causal Structures from Aviation Accident Narratives

    Compares rule-based NLP, transformer (DistilBERT), and few-shot LLM (Mistral-7B) approaches for extracting causal relationships from NTSB aviation accident narratives into knowledge graphs.

  • Energy-Efficient Avionics Security

    Research on energy-efficient security mechanisms for avionics systems.

  • Satellite Trajectory Projection and Collision Avoidance

    RL-based satellite collision avoidance environment and PPO pipeline.

  • Hierarchical Q-Learning for Threat Hunting (Dueling DQN)

    A multi-stage RL agent performing hierarchical intrusion detection over network traffic, including binary classification and attack-type prediction.

Research Posters

Below is a selected collection of research posters from 2022–2026, representing my work across reinforcement learning, optimization, sonar processing, remote sensing, sim-to-real robotics, and aviation NLP.

AERO-Sim2Real: Reinforcement Learning For Sim-to-Real Drone Deployment (2026)

Trains PPO-based flight-control policies in NVIDIA IsaacLab and deploys them on Crazyflie 2.1 nano-quadrotors, quantifying sim-to-real divergence across hover, waypoint-following, and figure-eight trajectory tasks.

Learning Causal Structures from Aviation Accident Narratives (2025)

Compares rule-based NLP, transformer (DistilBERT), and few-shot LLM (Mistral-7B) approaches for extracting causal relationships from NTSB aviation accident narratives.

Satellite Proposal Poster (2025)

Proposes a development framework for satellite anomaly detection, with its motivation, mission impact, and design plan.

Satellite Trajectory Research Poster (2024)

Full research results for PPO-based orbital collision avoidance, including stabilized low-fuel orbits and debris-avoidance maneuvers.

Hierarchical Dueling Q-Learning for Intrusion Detection (2024)

A two-agent Q-learning and Dueling DQN system for classifying and responding to cyberattacks, with near-perfect binary and multi-class accuracy on large intrusion datasets.

Sonar Contrastive Learning Research (2024)

A semi-supervised contrastive learning model for sonar-based object detection, with noise reduction strategies and adaptation potential for radar systems.

Autonomous Maritime Robotics Association Poster (2023)

Early work on underwater sonar-based detection, noise filtering strategies, and object tracking; the foundation of later remote-sensing and sonar research.

Aircraft Detection & Kalman Tracking Poster (2023)

Combines YOLO object detection with ORB keypoints and Kalman filtering for aircraft motion prediction. Presented at NCUR 2023; 1st Place at the Embry-Riddle Student Research Symposium.