Table of Contents
  1. Stanford Intelligent Systems Laboratory (SISL)
    1. Constrained POMDP planning using learned safety surrogates
    2. Belief-state planning for long-horizon POMDPs
    3. Survey of black-box safety validation algorithms
    4. Probabilistic risk assessment of autonomous vehicles
    5. Emergency evacuations under compounding levels of uncertainty
    6. Stress testing autonomous systems
    7. Surrogate model-based optimization
  2. Stanford Doerr School of Sustainability
    1. Sequential optimization of geothermal energy production
    2. Safe carbon capture and storage (CCS)
    3. Model-fidelity sensitivity analysis for POMDPs
    4. Sequential decision making for critical mineral exploration
  3. Xwing (Autonomous Cargo Aircraft)
    1. Failure probability estimation
    2. Certification of machine learning avionics systems
  4. NASA Ames Research Center
    1. Autonomous rover validation
  5. MIT Lincoln Laboratory
    1. Traffic advisory optimization
    2. Pilot response modeling
    3. Autonomous safety maneuvering
    4. Large-scale safety validation
    5. Algorithm specification languages
    6. Wildfire resource allocation
  6. Wentworth Institute of Technology
    1. Galactic rotation curve modeling

Stanford Intelligent Systems Laboratory (SISL)

My primary research is conducted as a graduate student researcher at SISL, advised by Mykel Kochenderfer.

Constrained POMDP planning using learned safety surrogates

Developing algorithms for safe planning in high-dimensional, long-horizon partially observable environments, applied to carbon capture and storage (CCS).

  1. Robert J. Moss, Arec Jamgochian, Johannes Fischer, Anthony Corso, and Mykel J. Kochenderfer, ConstrainedZero: Chance-Constrained POMDP Planning Using Learned Probabilistic Failure Surrogates and Adaptive Safety Constraints, International Joint Conference on Artificial Intelligence (IJCAI), 2024.

  2. Robert J. Moss, Arec Jamgochian, Johannes Fischer, Anthony Corso, and Mykel J. Kochenderfer, Chance-Constrained POMDP Planning with Learned Neural Network Surrogates, IJCAI Workshop on Trustworthy Interactive Decision-Making with Foundation Models, 2024.

Belief-state planning for long-horizon POMDPs

Developed algorithm to plan in high-dimensional, long-horizon POMDPs that learns approximately optimal surrogates of the value function and policy to replace heuristics in expensive Monte Carlo tree searches, applied to critical mineral exploration.

Robert J. Moss, Anthony Corso, Jef Caers, and Mykel J. Kochenderfer, BetaZero: Belief-State Planning for Long-Horizon POMDPs Using Learned Approximations, Reinforcement Learning Journal (RLJ), 2024.

Survey of black-box safety validation algorithms

Co-authored literature survey of algorithms for black-box safety validation of cyber-physical systems, analyzing and distilling over 100 publications.

Anthony Corso, Robert J. Moss, Mark Koren, Ritchie Lee, and Mykel J. Kochenderfer, A Survey of Algorithms for Black-Box Safety Validation of Cyber-Physical Systems, Journal of Artificial Intelligence Research (JAIR), 2021.

Probabilistic risk assessment of autonomous vehicles

Led the development of an autonomous vehicle risk assessment framework using risk analysis methods from the financial industry, demonstrated on 2D driving simulators and the 3D driving simulator CARLA [1]. Developed an algorithm to predict when failures will occur to increase failure rate of an autonomous vehicle [2] and developed an algorithm to detect weaknesses in autonomous vehicles [3].

  1. Robert J. Moss, Shubh Gupta, Robert Dyro, Karen Leung, Mykel J. Kochenderfer, Grace X. Gao, Marco Pavone, Edward Schmerling, Anthony Corso, Regina Madigan, Matei Stroila, and Tim Gibson, Autonomous Vehicle Risk Assessment, Stanford Center for AI Safety, 2021.

  2. Robert J. Moss, Predictive Risk for Efficient Black-Box Validation of Autonomous Vehicles, Stanford University (CS229: Machine Learning), 2021.

  3. Robert J. Moss, Adversarial Weakness Recognition for Efficient Black-Box Validation, Stanford University (CS230: Deep Learning), 2020.

Emergency evacuations under compounding levels of uncertainty

Developed a sequential decision support tool for emergency evacuation procedures and analyzed the sensitivity to different level of uncertainty. Demonstrated on a case study of the 2021 Afghanistan evacuation.

Lisa J. Einstein, Robert J. Moss, and Mykel J. Kochenderfer, Prioritizing Emergency Evacuations Under Compounding Levels of Uncertainty, IEEE Global Humanitairan Technology Conference (GHTC), 2022.

Stress testing autonomous systems

Developed reinforcement learning algorithm to stress test aircraft trajectory predictions in a flight management system (FMS) [1] implemented as an open-source package for general black-box systems [2]. Supported the use of the developed methods for certification of airborne software [3].

  1. Robert J. Moss, Lee Ritchie, Nicholas Visser, Joachim Hochwarth, James G. Lopez, and Mykel J. Kochenderfer, Adaptive Stress Testing of Trajectory Predictions in Flight Management Systems, AIAA/IEEE Digital Avionics Systems Conference (DASC), 2020.

  2. Robert J. Moss, POMDPStressTesting.jl: Adaptive Stress Testing for Black-Box Systems, Journal of Open Source Software (JOSS), 2021.

  3. Michael Durling, Heber Herencia-Zapana, Baoluo Meng, Mike Meiners, Joachim Hochwarth, Nicholas Visser, Ritchie Lee, Robert J. Moss, and Vidhya Tekken Valapil, Certification Considerations for Adaptive Stress Testing of Airborne Software, AIAA/IEEE Digital Avionics Systems Conference (DASC), 2021.

Surrogate model-based optimization

Developed stochastic optimization algorithms based on the cross-entropy method that use surrogates to find rare failure events in computationally expensive systems.

Robert J. Moss, Cross-Entropy Method Variants for Optimization, arXiv 2009.09043, 2020.

Stanford Doerr School of Sustainability

I'm also affiliated with the Stanford Center for Earth Resource Forecasting (SCERF) and the Stanford Mineral-X Initiative, working on safe geological planning problems.

Sequential optimization of geothermal energy production

Developing POMDP to optimize geothermal energy production based on a real-world case study dealing with large action spaces and small economical margins.

Research in progress.

Safe carbon capture and storage (CCS)

Supported the development of a POMDP for safe CO2_2 subsurface storage using a neural network surrogate model of CO2_2 plume migration dynamics.[1]
Random PolicyExpert PolicyPOMDP Policy

Yizheng Wang, Markus Zechner, Gege Wen, Anthony Louis Corso, John Michael Mern, Mykel J. Kochenderfer, and Jef Karel Caers, Optimizing Carbon Storage Operations for Long-Term Safety. arXiv 2304.09352 (Under Review), 2023.

[1] Figures courtesy of Yizheng Wang.

Model-fidelity sensitivity analysis for POMDPs

Studied the effect of model-fidelity on planning for critical mineral exploration and developed a general POMDP model-fidelity analysis framework [1,2].

  1. Robert J. Moss, Mariia Kozlova, Anthony Corso, and Jef Caers, Model-Fidelity Analysis for Sequential Decision-Making Systems Using Simulation Decomposition: Case Study of Critical Mineral Exploration, Routledge (Under Review), 2023.

  2. Mariia Kozlova, Robert J. Moss, Julian Scott Yeomans, and Jef Caers, Uncovering Heterogeneous Effects in Computational Models for Sustainable Decision-Making, Environmental Modelling & Software (Under Review), 2023.

Sequential decision making for critical mineral exploration

Supported the development of a belief-state MDP model of critical mineral exploration using reward shaping for more efficient and informed decision making.

Xwing (Autonomous Cargo Aircraft)

Failure probability estimation

Developed method using Bayesian optimization to falsify safety-critical subsystems of an autonomous aircraft and to estimate the probability of system failure using importance sampling. Applied to a neural network-based runway detection system in simulation.

  1. Robert J. Moss, Mykel J. Kochenderfer, Maxime Gariel, and Arthur Dubios, Bayesian Safety Validation for Failure Probability Estimation of Black-Box Systems, AIAA Journal of Aerospace Information Systems (JAIS), 2024.

  2. Robert J. Moss, Mykel J. Kochenderfer, Maxime Gariel, and Arthur Dubios, Bayesian Safety Validation for Black-Box Systems, AIAA AVIATION Forum, 2023.

Certification of machine learning avionics systems

Supported the development of formal and practical considerations for the certification of machine learning components in aircraft [3]. This research is currently being used to support FAA certification of an autonomous cargo aircraft.

Jean-Guillaume Durand, Arthur Dubois, and Robert J. Moss, Formal and Practical Elements for the Certification of Machine Learning Systems, AIAA/IEEE Digital Avionics Systems Conference (DASC), 2023.

NASA Ames Research Center

Autonomous rover validation

Applied safety validation techniques to stress test the decision making system of the autonomous lunar rover used in the VIPER mission, searching for water deposits on the Moon.[1]

[1] Figures courtesy of Edward Balaban.

MIT Lincoln Laboratory

Traffic advisory optimization

Researched global optimization methods to tune the traffic advisories (TAs) of the next-generation aircraft collision avoidance system (ACAS Xa) and developed a tool to efficiently optimize the TA logic over aircraft near-collision encounters from US and European airspace.

Pilot response modeling

Developed modeling technique using Bayesian networks that predicts the probability that an onboard pilot will respond to collision avoidance resolution advisory (RA) maneuvers [1, 2].

  1. Edward H. Londner, and Robert J. Moss, Bayesian Network Model of Pilot Response to Collision Avoidance System Resolution Advisories, Journal of Air Transportation (JAT), 2018.

  2. ———, A Bayesian Network Model of Pilot Response to TCAS Resolution Advisories, Air Traffic Management Research and Development Seminar (ATM R&D Seminar), 2017.

Autonomous safety maneuvering

Analyzed the safety of ACAS Xu on various vertical rate and turn rate limits which led the FAA to require all UAS equipped with ACAS Xu to autonomously respond to collision avoidance advisories.

Michael P. Owen, Adam Panken, Robert J. Moss, Luis Alvarez, and Charles Leeper, ACAS Xu: Integrated Collision Avoidance and Detect and Avoid Capability for UAS, AIAA/IEEE Digital Avionics Systems Conference (DASC), 2019.

Large-scale safety validation

Validated ACAS Xa and ACAS Xu on billions of realistic collision avoidance trajectories using parallel processing resources from the Lincoln Laboratory Supercomputing Center to support FAA certification of ACAS X. Developed a tool for document generation of safety and operational suitability metrics.

Adam Gjersvik, and Robert J. Moss, A Parallel Simulation Approach to ACAS X Development, IEEE High Performance Extreme Computing Conference (HPEC), 2019.

Algorithm specification languages

Studied the use of various programming languages to be used as a specification language for ACAS X to replace legacy pseudocode. Implemented ACAS X in the Julia programming language, enabling line-by-line evaluation of the published specification. Developed an automated tool to generate the 700+ page algorithm design document.

Robert J. Moss, Using Julia as a Specification Language for the Next-Generation Airborne Collision Avoidance System (ACAS X), JuliaCon, 2015.

Wildfire resource allocation

Developed a wildfire simulator and decision support tool for wildfire incident commanders to optimize the placement of aerial and land water resources to efficiently extinguish wildfires. Implemented realistic wildfire dynamics models and heuristics for better decisions using Monte Carlo tree search.

  1. J. Daniel Griffith, Mykel J. Kochenderfer, Robert J. Moss, Velibor V. Mišić, Vishal Gupta, and Dimitris Bertsimas, Automated Dynamic Resource Allocation for Wildfire Suppression, Lincoln Laboratory Journal, 2017.

  2. Dimitris Bertsimas, J. Daniel Griffith, Vishal Gupta, Mykel J. Kochenderfer, Velibor V. Mišić, and Robert J. Moss, A Comparison of Monte Carlo Tree Search and Mathematical Optimization for Large Scale Dynamic Resource, arXiv 1405.5498, 2014.

Wentworth Institute of Technology

Galactic rotation curve modeling

Developed modeling and simulation tool to study the fit of gravitational theories to observational data of galactic rotation. Implemented the χ2\chi^2 test for statistical hypothesis testing.

  1. James G. O'Brien, Thomas L. Chiarelli, Jeremy Dentico, Modestas Stulge, Brian Stefanski, Robert J. Moss, and Spasen and Chaykov, Alternative Gravity Rotation Curves for the LITTLE THINGS Survey, The Astrophysical Journal, 2018.

  2. James G. O'Brien, Spasen S. Chaykov, Jeremy Dentico, Modestas Stulge, Brian Stefanski, and Robert J. Moss, Recent Advancements in Conformal Gravity, Journal of Physics: Conference Series, 2017.

  3. James G. O'Brien, and Robert J. Moss, Rotation Curve for the Milky Way Galaxy in Conformal Gravity, Journal of Physics: Conference Series, 2015.

  4. Robert J. Moss, and James G. O'Brien, Rotation Curve Modeler: A Modeling and Simulation Tool for Arbitrary Galaxies, Wentworth Institute of Technology, 2014.