Mykel J. Kochenderfer

Max Lamparth, Daniel Fein, Andreas Haupt et al. · stanford

Single-axis mitigations of reward-model biases (e.g., reducing proxy reliance on length, sycophancy, or style) can rotate optimization pressure onto correlated proxies rather than eliminate it, a failure mode we call reward bias substitution. The failure is enabled by a measurement-versus-optimization gap between audit and policy-induced distributions during mitigation evaluation and policy training. We formalize mitigation outcomes into a regime taxonomy and prove that successful mitigation, bias substitution, and overcorrection produce identical observables under any audit-distribution scoring, including ranking accuracy and win-rate, even when granted oracle access to the true reward. Across published preference-learning mitigation work, no method we survey reports the evidence needed to certify successful mitigation. Augmenting evaluation with policy-induced distributions while tracking multiple biases provably closes the gap, and we translate this into actionable prescriptions for mitigation methods and benchmarks. We demonstrate bias substitution in language model RLHF, where a length penalty during GRPO training compresses responses as intended yet redirects optimization pressure onto confidence calibration, driving the policy into overconfidence while factual free-form accuracy falls. We also show a published length-debiasing operator that zeroes reward-length correlation on the audit distribution but reintroduces bias under best-of-N selection on three of four SOTA reward models, and a length-sycophancy coupling whose direction reverses under human-LLM judge disagreement.

Nicholas Rober, Sydney M. Katz, Chelsea Sidrane et al. · mit

As neural networks (NNs) become more prevalent in safety-critical applications such as control of vehicles, there is a growing need to certify that systems with NN components are safe. This paper presents a set of backward reachability approaches for safety certification of neural feedback loops (NFLs), i.e., closed-loop systems with NN control policies. While backward reachability strategies have been developed for systems without NN components, the nonlinearities in NN activation functions and general noninvertibility of NN weight matrices make backward reachability for NFLs a challenging problem. To avoid the difficulties associated with propagating sets backward through NNs, we introduce a framework that leverages standard forward NN analysis tools to efficiently find over-approximations to backprojection (BP) sets, i.e., sets of states for which an NN policy will lead a system to a given target set. We present frameworks for calculating BP over approximations for both linear and nonlinear systems with control policies represented by feedforward NNs and propose computationally efficient strategies. We use numerical results from a variety of models to showcase the proposed algorithms, including a demonstration of safety certification for a 6D system.

Kyle D. Julian, Mykel J. Kochenderfer

Deep neural networks can be trained to be efficient and effective controllers for dynamical systems; however, the mechanics of deep neural networks are complex and difficult to guarantee. This work presents a general approach for providing guarantees for deep neural network controllers over multiple time steps using a combination of reachability methods and open source neural network verification tools. By bounding the system dynamics and neural network outputs, the set of reachable states can be over-approximated to provide a guarantee that the system will never reach states outside the set. The method is demonstrated on the mountain car problem as well as an aircraft collision avoidance problem. Results show that this approach can provide neural network guarantees given a bounded dynamic model.

Joshua Ott, Edward Balaban, Mykel J. Kochenderfer

Adaptive Informative Path Planning with Multimodal Sensing (AIPPMS) considers the problem of an agent equipped with multiple sensors, each with different sensing accuracy and energy costs. The agent's goal is to explore the environment and gather information subject to its resource constraints in unknown, partially observable environments. Previous work has focused on the less general Adaptive Informative Path Planning (AIPP) problem, which considers only the effect of the agent's movement on received observations. The AIPPMS problem adds additional complexity by requiring that the agent reasons jointly about the effects of sensing and movement while balancing resource constraints with information objectives. We formulate the AIPPMS problem as a belief Markov decision process with Gaussian process beliefs and solve it using a sequential Bayesian optimization approach with online planning. Our approach consistently outperforms previous AIPPMS solutions by more than doubling the average reward received in almost every experiment while also reducing the root-mean-square error in the environment belief by 50%. We completely open-source our implementation to aid in further development and comparison.

Masha Itkina, Mykel J. Kochenderfer

Although neural networks have seen tremendous success as predictive models in a variety of domains, they can be overly confident in their predictions on out-of-distribution (OOD) data. To be viable for safety-critical applications, like autonomous vehicles, neural networks must accurately estimate their epistemic or model uncertainty, achieving a level of system self-awareness. Techniques for epistemic uncertainty quantification often require OOD data during training or multiple neural network forward passes during inference. These approaches may not be suitable for real-time performance on high-dimensional inputs. Furthermore, existing methods lack interpretability of the estimated uncertainty, which limits their usefulness both to engineers for further system development and to downstream modules in the autonomy stack. We propose the use of evidential deep learning to estimate the epistemic uncertainty over a low-dimensional, interpretable latent space in a trajectory prediction setting. We introduce an interpretable paradigm for trajectory prediction that distributes the uncertainty among the semantic concepts: past agent behavior, road structure, and social context. We validate our approach on real-world autonomous driving data, demonstrating superior performance over state-of-the-art baselines. Our code is available at: https://github.com/sisl/InterpretableSelfAwarePrediction.

Lisa J. Einstein, Robert J. Moss, Mykel J. Kochenderfer

Well-executed emergency evacuations can save lives and reduce suffering. However, decision makers struggle to determine optimal evacuation policies given the chaos, uncertainty, and value judgments inherent in emergency evacuations. We propose and analyze a decision support tool for pre-crisis training exercises for teams preparing for civilian evacuations and explore the tool in the case of the 2021 U.S.-led evacuation from Afghanistan. We use different classes of Markov decision processes (MDPs) to capture compounding levels of uncertainty in (1) the priority category of who appears next at the gate for evacuation, (2) the distribution of priority categories at the population level, and (3) individuals' claimed priority category. We compare the number of people evacuated by priority status under eight heuristic policies. The optimized MDP policy achieves the best performance compared to all heuristic baselines. We also show that accounting for the compounding levels of model uncertainty incurs added complexity without improvement in policy performance. Useful heuristics can be extracted from the optimized policies to inform human decision makers. We open-source all tools to encourage robust dialogue about the trade-offs, limitations, and potential of integrating algorithms into high-stakes humanitarian decision-making.

Kevin T. Carlberg, Antony Jameson, Mykel J. Kochenderfer et al.

Data I/O poses a significant bottleneck in large-scale CFD simulations; thus, practitioners would like to significantly reduce the number of times the solution is saved to disk, yet retain the ability to recover any field quantity (at any time instance) a posteriori. The objective of this work is therefore to accurately recover missing CFD data a posteriori at any time instance, given that the solution has been written to disk at only a relatively small number of time instances. We consider in particular high-order discretizations (e.g., discontinuous Galerkin), as such techniques are becoming increasingly popular for the simulation of highly separated flows. To satisfy this objective, this work proposes a methodology consisting of two stages: 1) dimensionality reduction and 2) dynamics learning. For dimensionality reduction, we propose a novel hierarchical approach. First, the method reduces the number of degrees of freedom within each element of the high-order discretization by applying autoencoders from deep learning. Second, the methodology applies principal component analysis to compress the global vector of encodings. This leads to a low-dimensional state, which associates with a nonlinear embedding of the original CFD data. For dynamics learning, we propose to apply regression techniques (e.g., kernel methods) to learn the discrete-time velocity characterizing the time evolution of this low-dimensional state. A numerical example on a large-scale CFD example characterized by nearly 13 million degrees of freedom illustrates the suitability of the proposed method in an industrial setting.

Jennifer She, Jayesh K. Gupta, Mykel J. Kochenderfer

Sparse and delayed rewards pose a challenge to single agent reinforcement learning. This challenge is amplified in multi-agent reinforcement learning (MARL) where credit assignment of these rewards needs to happen not only across time, but also across agents. We propose Agent-Time Attention (ATA), a neural network model with auxiliary losses for redistributing sparse and delayed rewards in collaborative MARL. We provide a simple example that demonstrates how providing agents with their own local redistributed rewards and shared global redistributed rewards motivate different policies. We extend several MiniGrid environments, specifically MultiRoom and DoorKey, to the multi-agent sparse delayed rewards setting. We demonstrate that ATA outperforms various baselines on many instances of these environments. Source code of the experiments is available at https://github.com/jshe/agent-time-attention.

Elysia Q. Smyers, Sydney M. Katz, Anthony L. Corso et al.

Designing robust machine learning systems remains an open problem, and there is a need for benchmark problems that cover both environmental changes and evaluation on a downstream task. In this work, we introduce AVOIDDS, a realistic object detection benchmark for the vision-based aircraft detect-and-avoid problem. We provide a labeled dataset consisting of 72,000 photorealistic images of intruder aircraft with various lighting conditions, weather conditions, relative geometries, and geographic locations. We also provide an interface that evaluates trained models on slices of this dataset to identify changes in performance with respect to changing environmental conditions. Finally, we implement a fully-integrated, closed-loop simulator of the vision-based detect-and-avoid problem to evaluate trained models with respect to the downstream collision avoidance task. This benchmark will enable further research in the design of robust machine learning systems for use in safety-critical applications. The AVOIDDS dataset and code are publicly available at https://purl.stanford.edu/hj293cv5980 and https://github.com/sisl/VisionBasedAircraftDAA respectively.

Michael H. Lim, Tyler J. Becker, Mykel J. Kochenderfer et al.

Partially observable Markov decision processes (POMDPs) provide a flexible representation for real-world decision and control problems. However, POMDPs are notoriously difficult to solve, especially when the state and observation spaces are continuous or hybrid, which is often the case for physical systems. While recent online sampling-based POMDP algorithms that plan with observation likelihood weighting have shown practical effectiveness, a general theory characterizing the approximation error of the particle filtering techniques that these algorithms use has not previously been proposed. Our main contribution is bounding the error between any POMDP and its corresponding finite sample particle belief MDP (PB-MDP) approximation. This fundamental bridge between PB-MDPs and POMDPs allows us to adapt any sampling-based MDP algorithm to a POMDP by solving the corresponding particle belief MDP, thereby extending the convergence guarantees of the MDP algorithm to the POMDP. Practically, this is implemented by using the particle filter belief transition model as the generative model for the MDP solver. While this requires access to the observation density model from the POMDP, it only increases the transition sampling complexity of the MDP solver by a factor of $\mathcal{O}(C)$, where $C$ is the number of particles. Thus, when combined with sparse sampling MDP algorithms, this approach can yield algorithms for POMDPs that have no direct theoretical dependence on the size of the state and observation spaces. In addition to our theoretical contribution, we perform five numerical experiments on benchmark POMDPs to demonstrate that a simple MDP algorithm adapted using PB-MDP approximation, Sparse-PFT, achieves performance competitive with other leading continuous observation POMDP solvers.

Zahra Shahrooei, Mykel J. Kochenderfer, Ali Baheri

Testing controllers in safety-critical systems is vital for ensuring their safety and preventing failures. In this paper, we address the falsification problem within learning-based closed-loop control systems through simulation. This problem involves the identification of counterexamples that violate system safety requirements and can be formulated as an optimization task based on these requirements. Using full-fidelity simulator data in this optimization problem can be computationally expensive. To improve efficiency, we propose a multi-fidelity Bayesian optimization falsification framework that harnesses simulators with varying levels of accuracy. Our proposed framework can transition between different simulators and establish meaningful relationships between them. Through multi-fidelity Bayesian optimization, we determine both the optimal system input likely to be a counterexample and the appropriate fidelity level for assessment. We evaluated our approach across various Gym environments, each featuring different levels of fidelity. Our experiments demonstrate that multi-fidelity Bayesian optimization is more computationally efficient than full-fidelity Bayesian optimization and other baseline methods in detecting counterexamples. A Python implementation of the algorithm is available at https://github.com/SAILRIT/MFBO_Falsification.

Kyle D. Julian, Shivam Sharma, Jean-Baptiste Jeannin et al.

The next generation of aircraft collision avoidance systems frame the problem as a Markov decision process and use dynamic programming to optimize the alerting logic. The resulting system uses a large lookup table to determine advisories given to pilots, but these tables can grow very large. To enable the system to operate on limited hardware, prior work investigated compressing the table using a deep neural network. However, ensuring that the neural network reliably issues safe advisories is important for certification. This work defines linearized regions where each advisory can be safely provided, allowing Reluplex, a neural network verification tool, to check if unsafe advisories are ever issued. A notional collision avoidance policy is generated and used to train a neural network representation. The neural networks are checked for unsafe advisories, resulting in the discovery of thousands of unsafe counterexamples.

Maneekwan Toyungyernsub, Esen Yel, Jiachen Li et al.

Detection and segmentation of moving obstacles, along with prediction of the future occupancy states of the local environment, are essential for autonomous vehicles to proactively make safe and informed decisions. In this paper, we propose a framework that integrates the two capabilities together using deep neural network architectures. Our method first detects and segments moving objects in the scene, and uses this information to predict the spatiotemporal evolution of the environment around autonomous vehicles. To address the problem of direct integration of both static-dynamic object segmentation and environment prediction models, we propose using occupancy-based environment representations across the whole framework. Our method is validated on the real-world Waymo Open Dataset and demonstrates higher prediction accuracy than baseline methods.

Bernard Lange, Jiachen Li, Mykel J. Kochenderfer

Navigating complex and dynamic environments requires autonomous vehicles (AVs) to reason about both visible and occluded regions. This involves predicting the future motion of observed agents, inferring occluded ones, and modeling their interactions based on vectorized scene representations of the partially observable environment. However, prior work on occlusion inference and trajectory prediction have developed in isolation, with the former based on simplified rasterized methods and the latter assuming full environment observability. We introduce the Scene Informer, a unified approach for predicting both observed agent trajectories and inferring occlusions in a partially observable setting. It uses a transformer to aggregate various input modalities and facilitate selective queries on occlusions that might intersect with the AV's planned path. The framework estimates occupancy probabilities and likely trajectories for occlusions, as well as forecast motion for observed agents. We explore common observability assumptions in both domains and their performance impact. Our approach outperforms existing methods in both occupancy prediction and trajectory prediction in partially observable setting on the Waymo Open Motion Dataset.

Joshua Ott, Sung-Kyun Kim, Amanda Bouman et al.

Robotic exploration of unknown environments is fundamentally a problem of decision making under uncertainty where the robot must account for uncertainty in sensor measurements, localization, action execution, as well as many other factors. For large-scale exploration applications, autonomous systems must overcome the challenges of sequentially deciding which areas of the environment are valuable to explore while safely evaluating the risks associated with obstacles and hazardous terrain. In this work, we propose a risk-aware meta-level decision making framework to balance the tradeoffs associated with local and global exploration. Meta-level decision making builds upon classical hierarchical coverage planners by switching between local and global policies with the overall objective of selecting the policy that is most likely to maximize reward in a stochastic environment. We use information about the environment history, traversability risk, and kinodynamic constraints to reason about the probability of successful policy execution to switch between local and global policies. We have validated our solution in both simulation and on a variety of large-scale real world hardware tests. Our results show that by balancing local and global exploration we are able to significantly explore large-scale environments more efficiently.

Kyle Hollins Wray, Stas Tiomkin, Mykel J. Kochenderfer et al.

Multi-objective optimization models that encode ordered sequential constraints provide a solution to model various challenging problems including encoding preferences, modeling a curriculum, and enforcing measures of safety. A recently developed theory of topological Markov decision processes (TMDPs) captures this range of problems for the case of discrete states and actions. In this work, we extend TMDPs towards continuous spaces and unknown transition dynamics by formulating, proving, and implementing the policy gradient theorem for TMDPs. This theoretical result enables the creation of TMDP learning algorithms that use function approximators, and can generalize existing deep reinforcement learning (DRL) approaches. Specifically, we present a new algorithm for a policy gradient in TMDPs by a simple extension of the proximal policy optimization (PPO) algorithm. We demonstrate this on a real-world multiple-objective navigation problem with an arbitrary ordering of objectives both in simulation and on a real robot.

Daniel Fein, Max Lamparth, Violet Xiang et al. · stanford

Reward Models (RMs) are crucial for online alignment of language models (LMs) with human preferences. However, RM-based preference-tuning is vulnerable to reward hacking, whereby LM policies learn undesirable behaviors from flawed RMs. By systematically measuring biases in five high-quality RMs, including the state-of-the-art, we find that issues persist despite prior work with respect to length, sycophancy, and overconfidence. We also discover new issues related to bias toward model-specific styles and answer-order. We categorize RM failures by complexity and propose a simple post-hoc intervention to mitigate low-complexity biases that arise from spurious correlations. Our proposed mechanistic reward shaping reduces targeted biases without degrading reward quality and while using minimal labeled data. The method is extensible to new biases, model-internal, and generalizes out-of-distribution.

Zahra Shahrooei, Mykel J. Kochenderfer, Ali Baheri

Simulation-based falsification is a practical testing method to increase confidence that the system will meet safety requirements. Because full-fidelity simulations can be computationally demanding, we investigate the use of simulators with different levels of fidelity. As a first step, we express the overall safety specification in terms of environmental parameters and structure this safety specification as an optimization problem. We propose a multi-fidelity falsification framework using Bayesian optimization, which is able to determine at which level of fidelity we should conduct a safety evaluation in addition to finding possible instances from the environment that cause the system to fail. This method allows us to automatically switch between inexpensive, inaccurate information from a low-fidelity simulator and expensive, accurate information from a high-fidelity simulator in a cost-effective way. Our experiments on various environments in simulation demonstrate that multi-fidelity Bayesian optimization has falsification performance comparable to single-fidelity Bayesian optimization but with much lower cost.

Kyle Brown, Dylan M. Asmar, Mac Schwager et al.

Mobile autonomous robots have the potential to revolutionize manufacturing processes. However, employing large robot fleets in manufacturing requires addressing challenges including collision-free movement in a shared workspace, effective multi-robot collaboration to manipulate and transport large payloads, complex task allocation due to coupled manufacturing processes, and spatial planning for parallel assembly and transportation of nested subassemblies. We propose a full algorithmic stack for large-scale multi-robot assembly planning that addresses these challenges and can synthesize construction plans for complex assemblies with thousands of parts in a matter of minutes. Our approach takes in a CAD-like product specification and automatically plans a full-stack assembly procedure for a group of robots to manufacture the product. We propose an algorithmic stack that comprises: (i) an iterative radial layout optimization procedure to define a global staging layout for the manufacturing facility, (ii) a graph-repair mixed-integer program formulation and a modified greedy task allocation algorithm to optimally allocate robots and robot sub-teams to assembly and transport tasks, (iii) a geometric heuristic and a hill-climbing algorithm to plan collaborative carrying configurations of robot sub-teams, and (iv) a distributed control policy that enables robots to execute the assembly motion plan collision-free. We also present an open-source multi-robot manufacturing simulator implemented in Julia as a resource to the research community, to test our algorithms and to facilitate multi-robot manufacturing research more broadly. Our empirical results demonstrate the scalability and effectiveness of our approach by generating plans to manufacture a LEGO model of a Saturn V launch vehicle with 1845 parts, 306 subassemblies, and 250 robots in under three minutes on a standard laptop computer.

Anthony L. Corso, Sydney M. Katz, Craig Innes et al.

Modern autonomous systems rely on perception modules to process complex sensor measurements into state estimates. These estimates are then passed to a controller, which uses them to make safety-critical decisions. It is therefore important that we design perception systems to minimize errors that reduce the overall safety of the system. We develop a risk-driven approach to designing perception systems that accounts for the effect of perceptual errors on the performance of the fully-integrated, closed-loop system. We formulate a risk function to quantify the effect of a given perceptual error on overall safety, and show how we can use it to design safer perception systems by including a risk-dependent term in the loss function and generating training data in risk-sensitive regions. We evaluate our techniques on a realistic vision-based aircraft detect and avoid application and show that risk-driven design reduces collision risk by 37% over a baseline system.

Jiachen Li, David Isele, Kanghoon Lee et al.

Deep reinforcement learning (DRL) provides a promising way for intelligent agents (e.g., autonomous vehicles) to learn to navigate complex scenarios. However, DRL with neural networks as function approximators is typically considered a black box with little explainability and often suffers from suboptimal performance, especially for autonomous navigation in highly interactive multi-agent environments. To address these issues, we propose three auxiliary tasks with spatio-temporal relational reasoning and integrate them into the standard DRL framework, which improves the decision making performance and provides explainable intermediate indicators. We propose to explicitly infer the internal states (i.e., traits and intentions) of surrounding agents (e.g., human drivers) as well as to predict their future trajectories in the situations with and without the ego agent through counterfactual reasoning. These auxiliary tasks provide additional supervision signals to infer the behavior patterns of other interactive agents. Multiple variants of framework integration strategies are compared. We also employ a spatio-temporal graph neural network to encode relations between dynamic entities, which enhances both internal state inference and decision making of the ego agent. Moreover, we propose an interactivity estimation mechanism based on the difference between predicted trajectories in these two situations, which indicates the degree of influence of the ego agent on other agents. To validate the proposed method, we design an intersection driving simulator based on the Intelligent Intersection Driver Model (IIDM) that simulates vehicles and pedestrians. Our approach achieves robust and state-of-the-art performance in terms of standard evaluation metrics and provides explainable intermediate indicators (i.e., internal states, and interactivity scores) for decision making.

Sydney M. Katz, Anthony L. Corso, Esen Yel et al.

Perception systems operate as a subcomponent of the general autonomy stack, and perception system designers often need to optimize performance characteristics while maintaining safety with respect to the overall closed-loop system. For this reason, it is useful to distill high-level safety requirements into component-level requirements on the perception system. In this work, we focus on efficiently determining sets of safe perception system performance characteristics given a black-box simulator of the fully-integrated, closed-loop system. We combine the advantages of common black-box estimation techniques such as Gaussian processes and threshold bandits to develop a new estimation method, which we call smoothing bandits. We demonstrate our method on a vision-based aircraft collision avoidance problem and show improvements in terms of both accuracy and efficiency over the Gaussian process and threshold bandit baselines.

Soyeon Jung, Amelia Hardy, Mykel J. Kochenderfer

Realistic aircraft trajectory models are useful in the design and validation of air traffic management (ATM) systems. Models of aircraft operated under instrument flight rules (IFR) require capturing the variability inherent in how aircraft follow standard flight procedures. The variability in aircraft behavior differs among flight stages. In this paper, we propose a simple probabilistic model that can learn this variability from procedural data and flight tracks collected from radar surveillance data. For each segment, we use a Gaussian mixture model to learn the deviations of aircraft trajectories from their procedures. Given new procedures, we generate synthetic trajectories by sampling a series of deviations from the Gaussian mixture model and reconstructing the aircraft trajectory using the deviations and the procedures. We extend this method to capture pairwise correlations between aircraft and show how a pairwise model can be used to generate traffic involving an arbitrary number of aircraft. We demonstrate the proposed models on the arrival tracks and procedures of the John F. Kennedy International Airport. Distributional similarity between the original and the synthetic trajectory dataset was evaluated using the Jensen-Shannon divergence between the empirical distributions of different variables and we provide qualitative analyses of the synthetic trajectories generated.

Dylan M. Asmar, Mykel J. Kochenderfer

The level of autonomy is increasing in systems spanning multiple domains, but these systems still experience failures. One way to mitigate the risk of failures is to integrate human oversight of the autonomous systems and rely on the human to take control when the autonomy fails. In this work, we formulate a method of collaborative decision making through action suggestions that improves action selection without taking control of the system. Our approach uses each suggestion efficiently by incorporating the implicit information shared through suggestions to modify the agent's belief and achieves better performance with fewer suggestions than naively following the suggested actions. We assume collaborative agents share the same objective and communicate through valid actions. By assuming the suggested action is dependent only on the state, we can incorporate the suggested action as an independent observation of the environment. The assumption of a collaborative environment enables us to use the agent's policy to estimate the distribution over action suggestions. We propose two methods that use suggested actions and demonstrate the approach through simulated experiments. The proposed methodology results in increased performance while also being robust to suboptimal suggestions.

Anthony Corso, Yizheng Wang, Markus Zechner et al.

Geological carbon capture and sequestration (CCS), where CO$_2$ is stored in subsurface formations, is a promising and scalable approach for reducing global emissions. However, if done incorrectly, it may lead to earthquakes and leakage of CO$_2$ back to the surface, harming both humans and the environment. These risks are exacerbated by the large amount of uncertainty in the structure of the storage formation. For these reasons, we propose that CCS operations be modeled as a partially observable Markov decision process (POMDP) and decisions be informed using automated planning algorithms. To this end, we develop a simplified model of CCS operations based on a 2D spillpoint analysis that retains many of the challenges and safety considerations of the real-world problem. We show how off-the-shelf POMDP solvers outperform expert baselines for safe CCS planning. This POMDP model can be used as a test bed to drive the development of novel decision-making algorithms for CCS operations.

Kanghoon Lee, Jiachen Li, David Isele et al.

Although deep reinforcement learning (DRL) has shown promising results for autonomous navigation in interactive traffic scenarios, existing work typically adopts a fixed behavior policy to control social vehicles in the training environment. This may cause the learned driving policy to overfit the environment, making it difficult to interact well with vehicles with different, unseen behaviors. In this work, we introduce an efficient method to train diverse driving policies for social vehicles as a single meta-policy. By randomizing the interaction-based reward functions of social vehicles, we can generate diverse objectives and efficiently train the meta-policy through guiding policies that achieve specific objectives. We further propose a training strategy to enhance the robustness of the ego vehicle's driving policy using the environment where social vehicles are controlled by the learned meta-policy. Our method successfully learns an ego driving policy that generalizes well to unseen situations with out-of-distribution (OOD) social agents' behaviors in a challenging uncontrolled T-intersection scenario.

Xiaobai Ma, David Isele, Jayesh K. Gupta et al.

Multi-agent reinforcement learning (MARL) provides an efficient way for simultaneously learning policies for multiple agents interacting with each other. However, in scenarios requiring complex interactions, existing algorithms can suffer from an inability to accurately anticipate the influence of self-actions on other agents. Incorporating an ability to reason about other agents' potential responses can allow an agent to formulate more effective strategies. This paper adopts a recursive reasoning model in a centralized-training-decentralized-execution framework to help learning agents better cooperate with or compete against others. The proposed algorithm, referred to as the Recursive Reasoning Graph (R2G), shows state-of-the-art performance on multiple multi-agent particle and robotics games.

Harrison Delecki, Masha Itkina, Bernard Lange et al.

Autonomous vehicles (AVs) rely on environment perception and behavior prediction to reason about agents in their surroundings. These perception systems must be robust to adverse weather such as rain, fog, and snow. However, validation of these systems is challenging due to their complexity and dependence on observation histories. This paper presents a method for characterizing failures of LiDAR-based perception systems for AVs in adverse weather conditions. We develop a methodology based in reinforcement learning to find likely failures in object tracking and trajectory prediction due to sequences of disturbances. We apply disturbances using a physics-based data augmentation technique for simulating LiDAR point clouds in adverse weather conditions. Experiments performed across a wide range of driving scenarios from a real-world driving dataset show that our proposed approach finds high likelihood failures with smaller input disturbances compared to baselines while remaining computationally tractable. Identified failures can inform future development of robust perception systems for AVs.

Patrick Slade, Zachary N. Sunberg, Mykel J. Kochenderfer

Real-world autonomous systems operate under uncertainty about both their pose and dynamics. Autonomous control systems must simultaneously perform estimation and control tasks to maintain robustness to changing dynamics or modeling errors. However, information gathering actions often conflict with optimal actions for reaching control objectives, requiring a trade-off between exploration and exploitation. The specific problem setting considered here is for discrete-time nonlinear systems, with process noise, input-constraints, and parameter uncertainty. This article frames this problem as a Bayes-adaptive Markov decision process and solves it online using Monte Carlo tree search with an unscented Kalman filter to account for process noise and parameter uncertainty. This method is compared with certainty equivalent model predictive control and a tree search method that approximates the QMDP solution, providing insight into when information gathering is useful. Discrete time simulations characterize performance over a range of process noise and bounds on unknown parameters. An offline optimization method is used to select the Monte Carlo tree search parameters without hand-tuning. In lieu of recursive feasibility guarantees, a probabilistic bounding heuristic is offered that increases the probability of keeping the state within a desired region.

Jiachen Li, Chuanbo Hua, Jinkyoo Park et al.

While the modeling of pair-wise relations has been widely studied in multi-agent interacting systems, its ability to capture higher-level and larger-scale group-wise activities is limited. In this paper, we propose a group-aware relational reasoning approach (named EvolveHypergraph) with explicit inference of the underlying dynamically evolving relational structures, and we demonstrate its effectiveness for multi-agent trajectory prediction. In addition to the edges between a pair of nodes (i.e., agents), we propose to infer hyperedges that adaptively connect multiple nodes to enable group-aware relational reasoning in an unsupervised manner without fixing the number of hyperedges. The proposed approach infers the dynamically evolving relation graphs and hypergraphs over time to capture the evolution of relations, which are used by the trajectory predictor to obtain future states. Moreover, we propose to regularize the smoothness of the relation evolution and the sparsity of the inferred graphs or hypergraphs, which effectively improves training stability and enhances the explainability of inferred relations. The proposed approach is validated on both synthetic crowd simulations and multiple real-world benchmark datasets. Our approach infers explainable, reasonable group-aware relations and achieves state-of-the-art performance in long-term prediction.

Yizheng Wang, Markus Zechner, Gege Wen et al.

To combat global warming and mitigate the risks associated with climate change, carbon capture and storage (CCS) has emerged as a crucial technology. However, safely sequestering CO2 in geological formations for long-term storage presents several challenges. In this study, we address these issues by modeling the decision-making process for carbon storage operations as a partially observable Markov decision process (POMDP). We solve the POMDP using belief state planning to optimize injector and monitoring well locations, with the goal of maximizing stored CO2 while maintaining safety. Empirical results in simulation demonstrate that our approach is effective in ensuring safe long-term carbon storage operations. We showcase the flexibility of our approach by introducing three different monitoring strategies and examining their impact on decision quality. Additionally, we introduce a neural network surrogate model for the POMDP decision-making process to handle the complex dynamics of the multi-phase flow. We also investigate the effects of different fidelity levels of the surrogate model on decision qualities.

Anthony Corso, David Karamadian, Romeo Valentin et al.

As machine learning (ML) systems increasingly permeate high-stakes settings such as healthcare, transportation, military, and national security, concerns regarding their reliability have emerged. Despite notable progress, the performance of these systems can significantly diminish due to adversarial attacks or environmental changes, leading to overconfident predictions, failures to detect input faults, and an inability to generalize in unexpected scenarios. This paper proposes a holistic assessment methodology for the reliability of ML systems. Our framework evaluates five key properties: in-distribution accuracy, distribution-shift robustness, adversarial robustness, calibration, and out-of-distribution detection. A reliability score is also introduced and used to assess the overall system reliability. To provide insights into the performance of different algorithmic approaches, we identify and categorize state-of-the-art techniques, then evaluate a selection on real-world tasks using our proposed reliability metrics and reliability score. Our analysis of over 500 models reveals that designing for one metric does not necessarily constrain others but certain algorithmic techniques can improve reliability across multiple metrics simultaneously. This study contributes to a more comprehensive understanding of ML reliability and provides a roadmap for future research and development.

Arec Jamgochian, Anthony Corso, Mykel J. Kochenderfer

Rather than augmenting rewards with penalties for undesired behavior, Constrained Partially Observable Markov Decision Processes (CPOMDPs) plan safely by imposing inviolable hard constraint value budgets. Previous work performing online planning for CPOMDPs has only been applied to discrete action and observation spaces. In this work, we propose algorithms for online CPOMDP planning for continuous state, action, and observation spaces by combining dual ascent with progressive widening. We empirically compare the effectiveness of our proposed algorithms on continuous CPOMDPs that model both toy and real-world safety-critical problems. Additionally, we compare against the use of online solvers for continuous unconstrained POMDPs that scalarize cost constraints into rewards, and investigate the effect of optimistic cost propagation.

Bernard Lange, Masha Itkina, Mykel J. Kochenderfer

Environment prediction frameworks are integral for autonomous vehicles, enabling safe navigation in dynamic environments. LiDAR generated occupancy grid maps (L-OGMs) offer a robust bird's eye-view scene representation that facilitates joint scene predictions without relying on manual labeling unlike commonly used trajectory prediction frameworks. Prior approaches have optimized deterministic L-OGM prediction architectures directly in grid cell space. While these methods have achieved some degree of success in prediction, they occasionally grapple with unrealistic and incorrect predictions. We claim that the quality and realism of the forecasted occupancy grids can be enhanced with the use of generative models. We propose a framework that decouples occupancy prediction into: representation learning and stochastic prediction within the learned latent space. Our approach allows for conditioning the model on other available sensor modalities such as RGB-cameras and high definition maps. We demonstrate that our approach achieves state-of-the-art performance and is readily transferable between different robotic platforms on the real-world NuScenes, Waymo Open, and a custom dataset we collected on an experimental vehicle platform.

Christopher Lazarus, Mykel J. Kochenderfer

Many approaches for verifying input-output properties of neural networks have been proposed recently. However, existing algorithms do not scale well to large networks. Recent work in the field of model compression studied binarized neural networks (BNNs), whose parameters and activations are binary. BNNs tend to exhibit a slight decrease in performance compared to their full-precision counterparts, but they can be easier to verify. This paper proposes a simple mixed integer programming formulation for BNN verification that leverages network structure. We demonstrate our approach by verifying properties of BNNs trained on the MNIST dataset and an aircraft collision avoidance controller. We compare the runtime of our approach against state-of-the-art verification algorithms for full-precision neural networks. The results suggest that the difficulty of training BNNs might be worth the reduction in runtime achieved by our verification algorithm.

Junyoung Park, Federico Berto, Arec Jamgochian et al.

In this paper, we propose Meta-SysId, a meta-learning approach to model sets of systems that have behavior governed by common but unknown laws and that differentiate themselves by their context. Inspired by classical modeling-and-identification approaches, Meta-SysId learns to represent the common law through shared parameters and relies on online optimization to compute system-specific context. Compared to optimization-based meta-learning methods, the separation between class parameters and context variables reduces the computational burden while allowing batch computations and a simple training scheme. We test Meta-SysId on polynomial regression, time-series prediction, model-based control, and real-world traffic prediction domains, empirically finding it outperforms or is competitive with meta-learning baselines.

Bernard Lange, Masha Itkina, Jiachen Li et al.

Environment prediction frameworks are critical for the safe navigation of autonomous vehicles (AVs) in dynamic settings. LiDAR-generated occupancy grid maps (L-OGMs) offer a robust bird's-eye view for the scene representation, enabling self-supervised joint scene predictions while exhibiting resilience to partial observability and perception detection failures. Prior approaches have focused on deterministic L-OGM prediction architectures within the grid cell space. While these methods have seen some success, they frequently produce unrealistic predictions and fail to capture the stochastic nature of the environment. Additionally, they do not effectively integrate additional sensor modalities present in AVs. Our proposed framework, Latent Occupancy Prediction (LOPR), performs stochastic L-OGM prediction in the latent space of a generative architecture and allows for conditioning on RGB cameras, maps, and planned trajectories. We decode predictions using either a single-step decoder, which provides high-quality predictions in real-time, or a diffusion-based batch decoder, which can further refine the decoded frames to address temporal consistency issues and reduce compression losses. Our experiments on the nuScenes and Waymo Open datasets show that all variants of our approach qualitatively and quantitatively outperform prior approaches.

I. Samuel Akinwande, Chelsea Sidrane, Mykel J. Kochenderfer et al.

As dynamical systems equipped with neural network controllers (neural feedback systems) become increasingly prevalent, it is critical to develop methods to ensure their safe operation. Verifying safety requires extending control theoretic analysis methods to these systems. Although existing techniques can efficiently handle linear neural feedback systems, relatively few scalable methods address the nonlinear case. We propose a novel algorithm for forward reachability analysis of nonlinear neural feedback systems. The approach leverages the structure of the nonlinear transition functions of the systems to compute tight polyhedral enclosures (i.e., abstractions). These enclosures, combined with the neural controller, are then encoded as a mixed-integer linear program (MILP). Optimizing this MILP yields a sound over-approximation of the forward-reachable set. Beyond the conference version of this work, we perform more extensive ablations, and introduce further optimizations to the algorithm. We evaluate our algorithm on representative benchmarks, and demonstrate significant improvements over the current state of the art.

Amelia F. Hardy, Houjun Liu, Allie Griffith et al.

Existing LLM red-teaming approaches prioritize high attack success rate, often resulting in high-perplexity prompts. This focus overlooks low-perplexity attacks that are more difficult to filter, more likely to arise during benign usage, and more impactful as negative downstream training examples. In response, we introduce ASTPrompter, a single-step optimization method that uses contrastive preference learning to train an attacker to maintain low perplexity while achieving a high attack success rate (ASR). ASTPrompter achieves an attack success rate 5.1 times higher on Llama-8.1B while using inputs that are 2.1 times more likely to occur according to the frozen LLM. Furthermore, our attack transfers to Mistral-7B, Qwen-7B, and TinyLlama in both black- and white-box settings. Lastly, by tuning a single hyperparameter in our method, we discover successful attack prefixes along an efficient frontier between ASR and perplexity, highlighting perplexity as a previously under-considered factor in red-teaming.

Christopher Lazarus, Mykel J. Kochenderfer

The use of neural networks as function approximators has enabled many advances in reinforcement learning (RL). The generalization power of neural networks combined with advances in RL algorithms has reignited the field of artificial intelligence. Despite their power, neural networks are considered black boxes, and their use in safety-critical settings remains a challenge. Recently, neural network verification has emerged as a way to certify safety properties of networks. Verification is a hard problem, and it is difficult to scale to large networks such as the ones used in deep reinforcement learning. We provide an approach to train RL policies that are more easily verifiable. We use binarized neural networks (BNNs), a type of network with mostly binary parameters. We present an RL algorithm tailored specifically for BNNs. After training BNNs for the Atari environments, we verify robustness properties.

Anthony Corso, Kyu-Young Kim, Shubh Gupta et al.

An important step in the design of autonomous systems is to evaluate the probability that a failure will occur. In safety-critical domains, the failure probability is extremely small so that the evaluation of a policy through Monte Carlo sampling is inefficient. Adaptive importance sampling approaches have been developed for rare event estimation but do not scale well to sequential systems with long horizons. In this work, we develop two adaptive importance sampling algorithms that can efficiently estimate the probability of rare events for sequential decision making systems. The basis for these algorithms is the minimization of the Kullback-Leibler divergence between a state-dependent proposal distribution and a target distribution over trajectories, but the resulting algorithms resemble policy gradient and value-based reinforcement learning. We apply multiple importance sampling to reduce the variance of our estimate and to address the issue of multi-modality in the optimal proposal distribution. We demonstrate our approach on a control task with both continuous and discrete actions spaces and show accuracy improvements over several baselines.

Romeo Valentin, Sydney M. Katz, Joonghyun Lee et al.

This paper addresses the challenge of probabilistic parameter estimation given measurement uncertainty in real-time. We provide a general formulation and apply this to pose estimation for an autonomous visual landing system. We present three probabilistic parameter estimators: a least-squares sampling approach, a linear approximation method, and a probabilistic programming estimator. To evaluate these estimators, we introduce novel closed-form expressions for measuring calibration and sharpness specifically for multivariate normal distributions. Our experimental study compares the three estimators under various noise conditions. We demonstrate that the linear approximation estimator can produce sharp and well-calibrated pose predictions significantly faster than the other methods but may yield overconfident predictions in certain scenarios. Additionally, we demonstrate that these estimators can be integrated with a Kalman filter for continuous pose estimation during a runway approach where we observe a 50\% improvement in sharpness while maintaining marginal calibration. This work contributes to the integration of data-driven computer vision models into complex safety-critical aircraft systems and provides a foundation for developing rigorous certification guidelines for such systems.

Arec Jamgochian, Hugo Buurmeijer, Kyle H. Wray et al.

Optimal plans in Constrained Partially Observable Markov Decision Processes (CPOMDPs) maximize reward objectives while satisfying hard cost constraints, generalizing safe planning under state and transition uncertainty. Unfortunately, online CPOMDP planning is extremely difficult in large or continuous problem domains. In many large robotic domains, hierarchical decomposition can simplify planning by using tools for low-level control given high-level action primitives (options). We introduce Constrained Options Belief Tree Search (COBeTS) to leverage this hierarchy and scale online search-based CPOMDP planning to large robotic problems. We show that if primitive option controllers are defined to satisfy assigned constraint budgets, then COBeTS will satisfy constraints anytime. Otherwise, COBeTS will guide the search towards a safe sequence of option primitives, and hierarchical monitoring can be used to achieve runtime safety. We demonstrate COBeTS in several safety-critical, constrained partially observable robotic domains, showing that it can plan successfully in continuous CPOMDPs while non-hierarchical baselines cannot.

Samuel I. Akinwande, Sydney M. Katz, Mykel J. Kochenderfer et al.

Forward reachability analysis is a dominant approach for verifying reach-avoid specifications in neural feedback systems, i.e., dynamical systems controlled by neural networks, and a number of directions have been proposed and studied. In contrast, far less attention has been given to backward reachability analysis for these systems, in part because of the limited scalability of known techniques. In this work, we begin to address this gap by introducing new algorithms for computing both over- and underapproximations of backward reachable sets for nonlinear neural feedback systems. We also describe and implement an integration of these backward reachability techniques with existing ones for forward analysis. We call the resulting algorithm Forward and Backward Reachability Integration for Certification (FaBRIC). We evaluate our algorithms on a representative set of benchmarks and show that they significantly outperform the prior state of the art.

Duncan Eddy, Esen Yel, Emma Passmore et al.

Managing announced task completion times is a fundamental control problem in project management. While extensive research exists on estimating task durations and task scheduling, the problem of when and how to update completion times communicated to stakeholders remains understudied. Organizations must balance announcement accuracy against the costs of frequent timeline updates, which can erode stakeholder trust and trigger costly replanning. Despite the prevalence of this problem, current approaches rely on static predictions or ad-hoc policies that fail to account for the sequential nature of announcement management. In this paper, we formulate the task announcement problem as a Partially Observable Markov Decision Process (POMDP) where the control policy must decide when to update announced completion times based on noisy observations of true task completion. Since most state variables (current time and previous announcements) are fully observable, we leverage the Mixed Observability MDP (MOMDP) framework to enable more efficient policy optimization. Our reward structure captures the dual costs of announcement errors and update frequency, enabling synthesis of optimal announcement control policies. Using off-the-shelf solvers, we generate policies that act as feedback controllers, adaptively managing announcements based on belief state evolution. Simulation results demonstrate significant improvements in both accuracy and announcement stability compared to baseline strategies, achieving up to 75\% reduction in unnecessary updates while maintaining or improving prediction accuracy.

Mahdi Al-Husseini, Mykel J. Kochenderfer, Kyle H. Wray

We introduce an expressive framework and algorithms for the semi-decentralized control of cooperative agents in environments with communication uncertainty. Whereas semi-Markov control admits a distribution over time for agent actions, semi-Markov communication, or what we refer to as semi-decentralization, gives a distribution over time for what actions and observations agents can store in their histories. We extend semi-decentralization to the partially observable Markov decision process (POMDP). The resulting SDec-POMDP unifies decentralized and multiagent POMDPs and several existing explicit communication mechanisms. We present recursive small-step semi-decentralized A* (RS-SDA*), an exact algorithm for generating optimal SDec-POMDP policies. RS-SDA* is evaluated on semi-decentralized versions of several standard benchmarks and a maritime medical evacuation scenario. This paper provides a well-defined theoretical foundation for exploring many classes of multiagent communication problems through the lens of semi-decentralization.

Gabriela Aránguiz Dias, Kiana Jafari, Allie Griffith et al.

Across healthcare, agentic artificial intelligence (AI) systems are increasingly promoted as capable of autonomous action, yet in practice they currently operate under near-total human oversight due to safety, regulatory, and liability constraints that make autonomous clinical reasoning infeasible in high-stakes environments. While market enthusiasm suggests a revolution in healthcare agents, the conceptual assumptions and accountability structures shaping these systems remain underexamined. We present a qualitative study based on interviews with 20 stakeholders, including developers, implementers, and end users. Our analysis identifies three mutually reinforcing tensions: conceptual fragmentation regarding the definition of `agentic'; an autonomy contradiction where commercial promises exceed operational reality; and an evaluation blind spot that prioritizes technical benchmarks over sociotechnical safety. We argue that agentic {AI} functions as a site of contested meaning-making where technical aspirations, commercial incentives, and clinical constraints intersect, carrying material consequences for patient safety and the distribution of blame.

Dylan M. Asmar, Mykel J. Kochenderfer

Autonomous agents operating in sequential decision-making tasks under uncertainty can benefit from external action suggestions, which provide valuable guidance but inherently vary in reliability. Existing methods for incorporating such advice typically assume static and known suggester quality parameters, limiting practical deployment. We introduce a framework that dynamically learns and adapts to varying suggester reliability in partially observable environments. First, we integrate suggester quality directly into the agent's belief representation, enabling agents to infer and adjust their reliance on suggestions through Bayesian inference over suggester types. Second, we introduce an explicit ``ask'' action allowing agents to strategically request suggestions at critical moments, balancing informational gains against acquisition costs. Experimental evaluation demonstrates robust performance across varying suggester qualities, adaptation to changing reliability, and strategic management of suggestion requests. This work provides a foundation for adaptive human-agent collaboration by addressing suggestion uncertainty in uncertain environments.

Marc R. Schlichting, Nina V. Boord, Anthony L. Corso et al.

Discovering potential failures of an autonomous system is important prior to deployment. Falsification-based methods are often used to assess the safety of such systems, but the cost of running many accurate simulation can be high. The validation can be accelerated by identifying critical failure scenarios for the system under test and by reducing the simulation runtime. We propose a Bayesian approach that integrates meta-learning strategies with a multi-armed bandit framework. Our method involves learning distributions over scenario parameters that are prone to triggering failures in the system under test, as well as a distribution over fidelity settings that enable fast and accurate simulations. In the spirit of meta-learning, we also assess whether the learned fidelity settings distribution facilitates faster learning of the scenario parameter distributions for new scenarios. We showcase our methodology using a cutting-edge 3D driving simulator, incorporating 16 fidelity settings for an autonomous vehicle stack that includes camera and lidar sensors. We evaluate various scenarios based on an autonomous vehicle pre-crash typology. As a result, our approach achieves a significant speedup, up to 18 times faster compared to traditional methods that solely rely on a high-fidelity simulator.

Romeo Valentin, Olivia Beyer Bruvik, Marc R. Schlichting et al.

EASA's learning-assurance guidance requires data-driven aviation systems to build and monitor their own situation representation, yet for neural networks the technical means to provide such evidence remain an open problem. We address this gap for a vision-based aircraft landing system: we propose that a minimally assurable model must at least be shown to separate content from style in its own situation representation. Showing that the model's predictions then rely largely on the contentful representation components leads to a concrete assurance path. To demonstrate this assurance path on a concrete model we train a vision transformer model for runway keypoint regression on the LARDv2 dataset. The model, which acts as the subject for our assurance demonstration, produces per-patch embeddings that we decompose into interpretable atoms via K-SVD sparse dictionary learning. A qualitative visualization confirms that contentful atoms track task-relevant runway structure and stylistic atoms track domain-specific appearance, and the regression head is shown to place almost all of its linear weight on contentful atoms. We further build on the content/style separation and define out-of-model-scope (OOMS) detection, a novel runtime assurance approach directly monitoring the model's situation representation. OOMS monitoring is complementary to operational design domain and output-space out-of-distribution monitoring and addresses concrete requirements of the recent EASA guidance. By directly analyzing a model's situation representation both at test time and runtime, this work delivers the first concrete piece of the representation-level evidence that EASA learning-assurance guidance demands, and points to mechanistic interpretability as a practical building block of future aviation safety cases.