Mykel Kochenderfer

Andreas Haupt, Justin Hartenstein, Anka Reuel et al.

AI benchmarks have well-documented limitations, with prior work examining contamination, saturation, and construct underspecification. Aggregation has received far less attention: benchmarks are typically summarized by uniformly averaging item-level scores, implicitly treating every test item as equally valuable. We model benchmarking as a multitask principal-agent game and show that the welfare loss from a benchmark is determined jointly by three item-level primitives: alignment with normative welfare priorities, marginal improvability, and performance variance. We translate the theory into an audit framework that ranks items along each of these three axes, and apply it to OLMES items using WORKBank for welfare, the EvoLM 4B suite for improvability, and the PolyPythias 410M panel for variance. The framework surfaces items that are Pareto-inferior within OLMES subject to a pro-worker welfare operationalization. All code is available at https://github.com/stair-lab/principal-agent-benchmarks.

Fan-Yun Sun, Isaac Kauvar, Ruohan Zhang et al. · stanford

Modeling multi-agent systems requires understanding how agents interact. Such systems are often difficult to model because they can involve a variety of types of interactions that layer together to drive rich social behavioral dynamics. Here we introduce a method for accurately modeling multi-agent systems. We present Interaction Modeling with Multiplex Attention (IMMA), a forward prediction model that uses a multiplex latent graph to represent multiple independent types of interactions and attention to account for relations of different strengths. We also introduce Progressive Layer Training, a training strategy for this architecture. We show that our approach outperforms state-of-the-art models in trajectory forecasting and relation inference, spanning three multi-agent scenarios: social navigation, cooperative task achievement, and team sports. We further demonstrate that our approach can improve zero-shot generalization and allows us to probe how different interactions impact agent behavior.

Mubashara Akhtar, Anka Reuel, Prajna Soni et al. · meta-ai

Artificial Intelligence (AI) benchmarks play a central role in measuring progress in model development and guiding deployment decisions. However, many benchmarks quickly become saturated, meaning that they can no longer differentiate between the best-performing models, diminishing their long-term value. In this study, we analyze benchmark saturation across 60 Large Language Model (LLM) benchmarks selected from technical reports by major model developers. To identify factors driving saturation, we characterize benchmarks along 14 properties spanning task design, data construction, and evaluation format. We test five hypotheses examining how each property contributes to saturation rates. Our analysis reveals that nearly half of the benchmarks exhibit saturation, with rates increasing as benchmarks age. Notably, hiding test data (i.e., public vs. private) shows no protective effect, while expert-curated benchmarks resist saturation better than crowdsourced ones. Our findings highlight which design choices extend benchmark longevity and inform strategies for more durable evaluation.

Bertrand Charpentier, Ransalu Senanayake, Mykel Kochenderfer et al.

Characterizing aleatoric and epistemic uncertainty on the predicted rewards can help in building reliable reinforcement learning (RL) systems. Aleatoric uncertainty results from the irreducible environment stochasticity leading to inherently risky states and actions. Epistemic uncertainty results from the limited information accumulated during learning to make informed decisions. Characterizing aleatoric and epistemic uncertainty can be used to speed up learning in a training environment, improve generalization to similar testing environments, and flag unfamiliar behavior in anomalous testing environments. In this work, we introduce a framework for disentangling aleatoric and epistemic uncertainty in RL. (1) We first define four desiderata that capture the desired behavior for aleatoric and epistemic uncertainty estimation in RL at both training and testing time. (2) We then present four RL models inspired by supervised learning (i.e. Monte Carlo dropout, ensemble, deep kernel learning models, and evidential networks) to instantiate aleatoric and epistemic uncertainty. Finally, (3) we propose a practical evaluation method to evaluate uncertainty estimation in model-free RL based on detection of out-of-distribution environments and generalization to perturbed environments. We present theoretical and experimental evidence to validate that carefully equipping model-free RL agents with supervised learning uncertainty methods can fulfill our desiderata.

Enna Sachdeva, Nakul Agarwal, Suhas Chundi et al.

The widespread adoption of commercial autonomous vehicles (AVs) and advanced driver assistance systems (ADAS) may largely depend on their acceptance by society, for which their perceived trustworthiness and interpretability to riders are crucial. In general, this task is challenging because modern autonomous systems software relies heavily on black-box artificial intelligence models. Towards this goal, this paper introduces a novel dataset, Rank2Tell, a multi-modal ego-centric dataset for Ranking the importance level and Telling the reason for the importance. Using various close and open-ended visual question answering, the dataset provides dense annotations of various semantic, spatial, temporal, and relational attributes of various important objects in complex traffic scenarios. The dense annotations and unique attributes of the dataset make it a valuable resource for researchers working on visual scene understanding and related fields. Furthermore, we introduce a joint model for joint importance level ranking and natural language captions generation to benchmark our dataset and demonstrate performance with quantitative evaluations.

Anka Reuel, Avijit Ghosh, Jenny Chim et al.

Foundation models are increasingly central to high-stakes AI systems, and governance frameworks now depend on evaluations to assess their risks and capabilities. Although general capability evaluations are widespread, social impact assessments covering bias, fairness, privacy, environmental costs, and labor practices remain uneven across the AI ecosystem. To characterize this landscape, we conduct the first comprehensive analysis of both first-party and third-party social impact evaluation reporting across a wide range of model developers. Our study examines 186 first-party release reports and 183 post-release evaluation sources, and complements this quantitative analysis with interviews of model developers. We find a clear division of evaluation labor: first-party reporting is sparse, often superficial, and has declined over time in key areas such as environmental impact and bias, while third-party evaluators including academic researchers, nonprofits, and independent organizations provide broader and more rigorous coverage of bias, harmful content, and performance disparities. However, this complementarity has limits. Only model developers can authoritatively report on data provenance, content moderation labor, financial costs, and training infrastructure, yet interviews reveal that these disclosures are often deprioritized unless tied to product adoption or regulatory compliance. Our findings indicate that current evaluation practices leave major gaps in assessing AI's societal impacts, highlighting the urgent need for policies that promote developer transparency, strengthen independent evaluation ecosystems, and create shared infrastructure to aggregate and compare third-party evaluations in a consistent and accessible way.

Kiana Jafari, Paul Ulrich Nikolaus Rust, Duncan Eddy et al.

Learning from human feedback~(LHF) assumes that expert judgments, appropriately aggregated, yield valid ground truth for training and evaluating AI systems. We tested this assumption in mental health, where high safety stakes make expert consensus essential. Three certified psychiatrists independently evaluated LLM-generated responses using a calibrated rubric. Despite similar training and shared instructions, inter-rater reliability was consistently poor ($ICC$ $0.087$--$0.295$), falling below thresholds considered acceptable for consequential assessment. Disagreement was highest on the most safety-critical items. Suicide and self-harm responses produced greater divergence than any other category, and was systematic rather than random. One factor yielded negative reliability (Krippendorff's $α= -0.203$), indicating structured disagreement worse than chance. Qualitative interviews revealed that disagreement reflects coherent but incompatible individual clinical frameworks, safety-first, engagement-centered, and culturally-informed orientations, rather than measurement error. By demonstrating that experts rely on holistic risk heuristics rather than granular factor discrimination, these findings suggest that aggregated labels function as arithmetic compromises that effectively erase grounded professional philosophies. Our results characterize expert disagreement in safety-critical AI as a sociotechnical phenomenon where professional experience introduces sophisticated layers of principled divergence. We discuss implications for reward modeling, safety classification, and evaluation benchmarks, recommending that practitioners shift from consensus-based aggregation to alignment methods that preserve and learn from expert disagreement.

Tina Diao, Samriddhi Singla, Ayan Mukhopadhyay et al.

Recent wildfires in the United States have resulted in loss of life and billions of dollars, destroying countless structures and forests. Fighting wildfires is extremely complex. It is difficult to observe the true state of fires due to smoke and risk associated with ground surveillance. There are limited resources to be deployed over a massive area and the spread of the fire is challenging to predict. This paper proposes a decision-theoretic approach to combat wildfires. We model the resource allocation problem as a partially-observable Markov decision process. We also present a data-driven model that lets us simulate how fires spread as a function of relevant covariates. A major problem in using data-driven models to combat wildfires is the lack of comprehensive data sources that relate fires with relevant covariates. We present an algorithmic approach based on large-scale raster and vector analysis that can be used to create such a dataset. Our data with over 2 million data points is the first open-source dataset that combines existing fire databases with covariates extracted from satellite imagery. Through experiments using real-world wildfire data, we demonstrate that our forecasting model can accurately model the spread of wildfires. Finally, we use simulations to demonstrate that our response strategy can significantly reduce response times compared to baseline methods.

Houjun Liu, Lisa Einstein, John Yang et al.

LLM coding agents now generate code at an unprecedented scale, yet LLM-generated code introduces cybersecurity vulnerabilities into codebases without human involvement. Even when frontier models are explicitly asked to write secure production code with relevant weaknesses to avoid in context, we find that they still produce verifiable vulnerabilities on average 23% of the time across a corpus of 250 benign coding prompts. We introduce SecureForge, an automated pipeline that both audits security risks of frontier models and produces auditing-informed secure system prompts that reduce output security vulnerabilities while maintaining unit test performance. SecureForge first identifies benign prompts that produce statically detectable vulnerabilities, and then amplifies them into a large synthetic prompt corpus of diverse scenarios using a Markovian sampling technique to jointly maintain error rates and prompt diversity. This corpus is then used to iteratively optimize the system prompts to reduce output security vulnerabilities. On frontier models, SecureForge yields a statistically significant Pareto improvement in both unit test success and output security, with output vulnerabilities reduced by up to 48%. The resulting system prompts transfer zero-shot to in-the-wild coding agent prompts, without any exposure to real user prompt distributions during optimization.

Francois Chaubard, Mykel Kochenderfer

During inference, Recurrent Neural Networks (RNNs) scale constant in both FLOPs and GPU memory with increasing context length, as they compress all prior tokens into a fixed-size memory. In contrast, transformers scale linearly in FLOPs and, at best, linearly in memory during generation, since they must attend to all previous tokens explicitly. Despite this inference-time advantage, training large RNNs on long contexts remains impractical because standard optimization methods depend on Backpropagation Through Time (BPTT). BPTT requires retention of all intermediate activations during the forward pass, causing memory usage to scale linearly with both context length and model size. In this paper, we show that Zero-Order Optimization (ZOO) methods such as Random-vector Gradient Estimation (RGE) can successfully replace BPTT to train RNNs with convergence rates that match, or exceed BPTT by up to 19 fold, while using orders of magnitude less memory and cost, as the model remains in inference mode throughout training. We further demonstrate that Central-Difference RGE (CD-RGE) corresponds to optimizing a smoothed surrogate loss, inherently regularizing training and improving generalization. Our method matches or outperforms BPTT across three settings: (1) overfitting, (2) transduction, and (3) language modeling. Across all tasks, with sufficient perturbations, our models generalize as well as or better than those trained with BPTT, often in fewer steps. Despite the need for more forward passes per step, we can surpass BPTT wall-clock time per step using recent advancements such as FlashRNN and distributed inference.

Bernard Lange, Anil Yildiz, Mansur Arief et al.

Developing general-purpose navigation policies for unknown environments remains a core challenge in robotics. Most existing systems rely on task-specific neural networks and fixed information flows, limiting their generalizability. Large Vision-Language Models (LVLMs) offer a promising alternative by embedding human-like knowledge for reasoning and planning, but prior LVLM-robot integrations have largely depended on pre-mapped spaces, hard-coded representations, and rigid control logic. We introduce the Agentic Robotic Navigation Architecture (ARNA), a general-purpose framework that equips an LVLM-based agent with a library of perception, reasoning, and navigation tools drawn from modern robotic stacks. At runtime, the agent autonomously defines and executes task-specific workflows that iteratively query modules, reason over multimodal inputs, and select navigation actions. This agentic formulation enables robust navigation and reasoning in previously unmapped environments, offering a new perspective on robotic stack design. Evaluated in Habitat Lab on the HM-EQA benchmark, ARNA outperforms state-of-the-art EQA-specific approaches. Qualitative results on RxR and custom tasks further demonstrate its ability to generalize across a broad range of navigation challenges.

Kiana Jafari Meimandi, Anka Reuel, Gabriela Aranguiz-Dias et al.

This paper examines the assessment challenges of Responsible AI (RAI) governance efforts in globally decentralized organizations through a case study collaboration between a leading research university and a multinational enterprise. While there are many proposed frameworks for RAI, their application in complex organizational settings with distributed decision-making authority remains underexplored. Our RAI assessment, conducted across multiple business units and AI use cases, reveals four key patterns that shape RAI implementation: (1) complex interplay between group-level guidance and local interpretation, (2) challenges translating abstract principles into operational practices, (3) regional and functional variation in implementation approaches, and (4) inconsistent accountability in risk oversight. Based on these findings, we propose an Adaptive RAI Governance (ARGO) Framework that balances central coordination with local autonomy through three interdependent layers: shared foundation standards, central advisory resources, and contextual local implementation. We contribute insights from academic-industry collaboration for RAI assessments, highlighting the importance of modular governance approaches that accommodate organizational complexity while maintaining alignment with responsible AI principles. These lessons offer practical guidance for organizations navigating the transition from RAI principles to operational practice within decentralized structures.

Gabriel Maher, Stephen Boyd, Mykel Kochenderfer et al.

We describe a light-weight yet performant system for hyper-parameter optimization that approximately minimizes an overall scalar cost function that is obtained by combining multiple performance objectives using a target-priority-limit scalarizer. It also supports a trade-off mode, where the goal is to find an appropriate trade-off among objectives by interacting with the user. We focus on the common scenario where there are on the order of tens of hyper-parameters, each with various attributes such as a range of continuous values, or a finite list of values, and whether it should be treated on a linear or logarithmic scale. The system supports multiple asynchronous simulations and is robust to simulation stragglers and failures.

Shushman Choudhury, Kiril Solovey, Mykel Kochenderfer et al.

We address the problem of routing a team of drones and trucks over large-scale urban road networks. To conserve their limited flight energy, drones can use trucks as temporary modes of transit en route to their own destinations. Such coordination can yield significant savings in total vehicle distance traveled, i.e., truck travel distance and drone flight distance, compared to operating drones and trucks independently. But it comes at the potentially prohibitive computational cost of deciding which trucks and drones should coordinate and when and where it is most beneficial to do so. We tackle this fundamental trade-off by decoupling our overall intractable problem into tractable sub-problems that we solve stage-wise. The first stage solves only for trucks, by computing paths that make them more likely to be useful transit options for drones. The second stage solves only for drones, by routing them over a composite of the road network and the transit network defined by truck paths from the first stage. We design a comprehensive algorithmic framework that frames each stage as a multi-agent path-finding problem and implement two distinct methods for solving them. We evaluate our approach on extensive simulations with up to $100$ agents on the real-world Manhattan road network containing nearly $4500$ vertices and $10000$ edges. Our framework saves on more than $50\%$ of vehicle distance traveled compared to independently solving for trucks and drones, and computes solutions for all settings within $5$ minutes on commodity hardware.

Raunak Bhattacharyya, Soyeon Jung, Liam Kruse et al.

Autonomous vehicles need to model the behavior of surrounding human driven vehicles to be safe and efficient traffic participants. Existing approaches to modeling human driving behavior have relied on both data-driven and rule-based methods. While data-driven models are more expressive, rule-based models are interpretable, which is an important requirement for safety-critical domains like driving. However, rule-based models are not sufficiently representative of data, and data-driven models are yet unable to generate realistic traffic simulation due to unrealistic driving behavior such as collisions. In this paper, we propose a methodology that combines rule-based modeling with data-driven learning. While the rules are governed by interpretable parameters of the driver model, these parameters are learned online from driving demonstration data using particle filtering. We perform driver modeling experiments on the task of highway driving and merging using data from three real-world driving demonstration datasets. Our results show that driver models based on our hybrid rule-based and data-driven approach can accurately capture real-world driving behavior. Further, we assess the realism of the driving behavior generated by our model by having humans perform a driving Turing test, where they are asked to distinguish between videos of real driving and those generated using our driver models.

Geoffrey Pettet, Ayan Mukhopadhyay, Mykel Kochenderfer et al.

A classical problem in city-scale cyber-physical systems (CPS) is resource allocation under uncertainty. Typically, such problems are modeled as Markov (or semi-Markov) decision processes. While online, offline, and decentralized approaches have been applied to such problems, they have difficulty scaling to large decision problems. We present a general approach to hierarchical planning that leverages structure in city-level CPS problems for resource allocation under uncertainty. We use the emergency response as a case study and show how a large resource allocation problem can be split into smaller problems. We then create a principled framework for solving the smaller problems and tackling the interaction between them. Finally, we use real-world data from Nashville, Tennessee, a major metropolitan area in the United States, to validate our approach. Our experiments show that the proposed approach outperforms state-of-the-art approaches used in the field of emergency response.

Kyle Hatch, John Mern, Mykel Kochenderfer

A collision avoidance system based on simple digital cameras would help enable the safe integration of small UAVs into crowded, low-altitude environments. In this work, we present an obstacle avoidance system for small UAVs that uses a monocular camera with a hybrid neural network and path planner controller. The system is comprised of a vision network for estimating depth from camera images, a high-level control network, a collision prediction network, and a contingency policy. This system is evaluated on a simulated UAV navigating an obstacle course in a constrained flight pattern. Results show the proposed system achieves low collision rates while maintaining operationally relevant flight speeds.

Jiaxuan You, Xiaobai Ma, Daisy Yi Ding et al.

Machine learning with missing data has been approached in two different ways, including feature imputation where missing feature values are estimated based on observed values, and label prediction where downstream labels are learned directly from incomplete data. However, existing imputation models tend to have strong prior assumptions and cannot learn from downstream tasks, while models targeting label prediction often involve heuristics and can encounter scalability issues. Here we propose GRAPE, a graph-based framework for feature imputation as well as label prediction. GRAPE tackles the missing data problem using a graph representation, where the observations and features are viewed as two types of nodes in a bipartite graph, and the observed feature values as edges. Under the GRAPE framework, the feature imputation is formulated as an edge-level prediction task and the label prediction as a node-level prediction task. These tasks are then solved with Graph Neural Networks. Experimental results on nine benchmark datasets show that GRAPE yields 20% lower mean absolute error for imputation tasks and 10% lower for label prediction tasks, compared with existing state-of-the-art methods.

Ayan Mukhopadhyay, Geoffrey Pettet, Mykel Kochenderfer et al.

Emergency response to incidents such as accidents, crimes, and fires is a major problem faced by communities. Emergency response management comprises of several stages and sub-problems like forecasting, resource allocation, and dispatch. The design of principled approaches to tackle each problem is necessary to create efficient emergency response management (ERM) pipelines. Over the last six years, we have worked with several first responder organizations to design ERM pipelines. In this paper, we highlight some of the challenges that we have identified and lessons that we have learned through our experience in this domain. Such challenges are particularly relevant for practitioners and researchers, and are important considerations even in the design of response strategies to mitigate disasters like floods and earthquakes.

Raunak Bhattacharyya, Blake Wulfe, Derek Phillips et al.

An open problem in autonomous vehicle safety validation is building reliable models of human driving behavior in simulation. This work presents an approach to learn neural driving policies from real world driving demonstration data. We model human driving as a sequential decision making problem that is characterized by non-linearity and stochasticity, and unknown underlying cost functions. Imitation learning is an approach for generating intelligent behavior when the cost function is unknown or difficult to specify. Building upon work in inverse reinforcement learning (IRL), Generative Adversarial Imitation Learning (GAIL) aims to provide effective imitation even for problems with large or continuous state and action spaces, such as modeling human driving. This article describes the use of GAIL for learning-based driver modeling. Because driver modeling is inherently a multi-agent problem, where the interaction between agents needs to be modeled, this paper describes a parameter-sharing extension of GAIL called PS-GAIL to tackle multi-agent driver modeling. In addition, GAIL is domain agnostic, making it difficult to encode specific knowledge relevant to driving in the learning process. This paper describes Reward Augmented Imitation Learning (RAIL), which modifies the reward signal to provide domain-specific knowledge to the agent. Finally, human demonstrations are dependent upon latent factors that may not be captured by GAIL. This paper describes Burn-InfoGAIL, which allows for disentanglement of latent variability in demonstrations. Imitation learning experiments are performed using NGSIM, a real-world highway driving dataset. Experiments show that these modifications to GAIL can successfully model highway driving behavior, accurately replicating human demonstrations and generating realistic, emergent behavior in the traffic flow arising from the interaction between driving agents.

Ayan Mukhopadhyay, Geoffrey Pettet, Sayyed Vazirizade et al.

In the last fifty years, researchers have developed statistical, data-driven, analytical, and algorithmic approaches for designing and improving emergency response management (ERM) systems. The problem has been noted as inherently difficult and constitutes spatio-temporal decision making under uncertainty, which has been addressed in the literature with varying assumptions and approaches. This survey provides a detailed review of these approaches, focusing on the key challenges and issues regarding four sub-processes: (a) incident prediction, (b) incident detection, (c) resource allocation, and (c) computer-aided dispatch for emergency response. We highlight the strengths and weaknesses of prior work in this domain and explore the similarities and differences between different modeling paradigms. We conclude by illustrating open challenges and opportunities for future research in this complex domain.

Zachary Sunberg, Mykel Kochenderfer

This work examines the hypothesis that partially observable Markov decision process (POMDP) planning with human driver internal states can significantly improve both safety and efficiency in autonomous freeway driving. We evaluate this hypothesis in a simulated scenario where an autonomous car must safely perform three lane changes in rapid succession. Approximate POMDP solutions are obtained through the partially observable Monte Carlo planning with observation widening (POMCPOW) algorithm. This approach outperforms over-confident and conservative MDP baselines and matches or outperforms QMDP. Relative to the MDP baselines, POMCPOW typically cuts the rate of unsafe situations in half or increases the success rate by 50%.

Raunak Bhattacharyya, Ransalu Senanayake, Kyle Brown et al.

Driver models are invaluable for planning in autonomous vehicles as well as validating their safety in simulation. Highly parameterized black-box driver models are very expressive, and can capture nuanced behavior. However, they usually lack interpretability and sometimes exhibit unrealistic-even dangerous-behavior. Rule-based models are interpretable, and can be designed to guarantee "safe" behavior, but are less expressive due to their low number of parameters. In this article, we show that online parameter estimation applied to the Intelligent Driver Model captures nuanced individual driving behavior while providing collision free trajectories. We solve the online parameter estimation problem using particle filtering, and benchmark performance against rule-based and black-box driver models on two real world driving data sets. We evaluate the closeness of our driver model to ground truth data demonstration and also assess the safety of the resulting emergent driving behavior.

Andrea Zanette, Alessandro Lazaric, Mykel Kochenderfer et al.

We study the exploration problem with approximate linear action-value functions in episodic reinforcement learning under the notion of low inherent Bellman error, a condition normally employed to show convergence of approximate value iteration. First we relate this condition to other common frameworks and show that it is strictly more general than the low rank (or linear) MDP assumption of prior work. Second we provide an algorithm with a high probability regret bound $\widetilde O(\sum_{t=1}^H d_t \sqrt{K} + \sum_{t=1}^H \sqrt{d_t} \IBE K)$ where $H$ is the horizon, $K$ is the number of episodes, $\IBE$ is the value if the inherent Bellman error and $d_t$ is the feature dimension at timestep $t$. In addition, we show that the result is unimprovable beyond constants and logs by showing a matching lower bound. This has two important consequences: 1) it shows that exploration is possible using only \emph{batch assumptions} with an algorithm that achieves the optimal statistical rate for the setting we consider, which is more general than prior work on low-rank MDPs 2) the lack of closedness (measured by the inherent Bellman error) is only amplified by $\sqrt{d_t}$ despite working in the online setting. Finally, the algorithm reduces to the celebrated \textsc{LinUCB} when $H=1$ but with a different choice of the exploration parameter that allows handling misspecified contextual linear bandits. While computational tractability questions remain open for the MDP setting, this enriches the class of MDPs with a linear representation for the action-value function where statistically efficient reinforcement learning is possible.

Geoffrey Pettet, Ayan Mukhopadhyay, Mykel Kochenderfer et al.

Emergency Response Management (ERM) is a critical problem faced by communities across the globe. Despite this, it is common for ERM systems to follow myopic decision policies in the real world. Principled approaches to aid ERM decision-making under uncertainty have been explored but have failed to be accepted into real systems. We identify a key issue impeding their adoption --- algorithmic approaches to emergency response focus on reactive, post-incident dispatching actions, i.e. optimally dispatching a responder \textit{after} incidents occur. However, the critical nature of emergency response dictates that when an incident occurs, first responders always dispatch the closest available responder to the incident. We argue that the crucial period of planning for ERM systems is not post-incident, but between incidents. This is not a trivial planning problem --- a major challenge with dynamically balancing the spatial distribution of responders is the complexity of the problem. An orthogonal problem in ERM systems is planning under limited communication, which is particularly important in disaster scenarios that affect communication networks. We address both problems by proposing two partially decentralized multi-agent planning algorithms that utilize heuristics and exploit the structure of the dispatch problem. We evaluate our proposed approach using real-world data, and find that in several contexts, dynamic re-balancing the spatial distribution of emergency responders reduces both the average response time as well as its variance.

Sheng Li, Maxim Egorov, Mykel Kochenderfer

New methodologies will be needed to ensure the airspace remains safe and efficient as traffic densities rise to accommodate new unmanned operations. This paper explores how unmanned free-flight traffic may operate in dense airspace. We develop and analyze autonomous collision avoidance systems for aircraft operating in dense airspace where traditional collision avoidance systems fail. We propose a metric for quantifying the decision burden on a collision avoidance system as well as a metric for measuring the impact of the collision avoidance system on airspace. We use deep reinforcement learning to compute corrections for an existing collision avoidance approach to account for dense airspace. The results show that a corrected collision avoidance system can operate more efficiently than traditional methods in dense airspace while maintaining high levels of safety.

Mark Koren, Mykel Kochenderfer

During the development of autonomous systems such as driverless cars, it is important to characterize the scenarios that are most likely to result in failure. Adaptive Stress Testing (AST) provides a way to search for the most-likely failure scenario as a Markov decision process (MDP). Our previous work used a deep reinforcement learning (DRL) solver to identify likely failure scenarios. However, the solver's use of a feed-forward neural network with a discretized space of possible initial conditions poses two major problems. First, the system is not treated as a black box, in that it requires analyzing the internal state of the system, which leads to considerable implementation complexities. Second, in order to simulate realistic settings, a new instance of the solver needs to be run for each initial condition. Running a new solver for each initial condition not only significantly increases the computational complexity, but also disregards the underlying relationship between similar initial conditions. We provide a solution to both problems by employing a recurrent neural network that takes a set of initial conditions from a continuous space as input. This approach enables robust and efficient detection of failures because the solution generalizes across the entire space of initial conditions. By simulating an instance where an autonomous car drives while a pedestrian is crossing a road, we demonstrate the solver is now capable of finding solutions for problems that would have previously been intractable.

John Mern, Dorsa Sadigh, Mykel Kochenderfer

Although deep reinforcement learning has advanced significantly over the past several years, sample efficiency remains a major challenge. Careful choice of input representations can help improve efficiency depending on the structure present in the problem. In this work, we present an attention-based method to project inputs into an efficient representation space that is invariant under changes to input ordering. We show that our proposed representation results in a search space that is a factor of m! smaller for inputs of m objects. Our experiments demonstrate improvements in sample efficiency for policy gradient methods on a variety of tasks. We show that our representation allows us to solve problems that are otherwise intractable when using naive approaches.

Duncan Eddy, Mykel Kochenderfer

Small satellites have proven to be viable Earth observation platforms. These satellites operate in regimes of increased trajectory uncertainty where traditional planning approaches can lead to sub-optimal task plans, limiting science return. Previous formulations of the space mission planning problem decouple trajectory prediction and planning, which leads to task plans that are less robust to uncertainty. We present a Markov decision process formulation of the problem that accounts for uncertainties by incorporating a distribution of possible collection windows characterized through Monte Carlo simulation. An approximate solution technique yields tasking schedules with rewards comparable to the conventional methods while simultaneously reducing the variations caused by uncertainties and improving runtime.

Shane Barratt, Mykel Kochenderfer, Stephen Boyd

Models for predicting aircraft motion are an important component of modern aeronautical systems. These models help aircraft plan collision avoidance maneuvers and help conduct offline performance and safety analyses. In this article, we develop a method for learning a probabilistic generative model of aircraft motion in terminal airspace, the controlled airspace surrounding a given airport. The method fits the model based on a historical dataset of radar-based position measurements of aircraft landings and takeoffs at that airport. We find that the model generates realistic trajectories, provides accurate predictions, and captures the statistical properties of aircraft trajectories. Furthermore, the model trains quickly, is compact, and allows for efficient real-time inference.

Zachary Sunberg, Mykel Kochenderfer

Online solvers for partially observable Markov decision processes have been applied to problems with large discrete state spaces, but continuous state, action, and observation spaces remain a challenge. This paper begins by investigating double progressive widening (DPW) as a solution to this challenge. However, we prove that this modification alone is not sufficient because the belief representations in the search tree collapse to a single particle causing the algorithm to converge to a policy that is suboptimal regardless of the computation time. This paper proposes and evaluates two new algorithms, POMCPOW and PFT-DPW, that overcome this deficiency by using weighted particle filtering. Simulation results show that these modifications allow the algorithms to be successful where previous approaches fail.

John Mern, Jayesh K Gupta, Mykel Kochenderfer

Deep artificial neural networks (ANNs) can represent a wide range of complex functions. Implementing ANNs in Von Neumann computing systems, though, incurs a high energy cost due to the bottleneck created between CPU and memory. Implementation on neuromorphic systems may help to reduce energy demand. Conventional ANNs must be converted into equivalent Spiking Neural Networks (SNNs) in order to be deployed on neuromorphic chips. This paper presents a way to perform this translation. We map the ANN weights to SNN synapses layer-by-layer by forming a least-square-error approximation problem at each layer. An optimal set of synapse weights may then be found for a given choice of ANN activation function and SNN neuron. Using an appropriate constrained solver, we can generate SNNs compatible with digital, analog, or hybrid chip architectures. We present an optimal node pruning method to allow SNN layer sizes to be set by the designer. To illustrate this process, we convert three ANNs, including one convolutional network, to SNNs. In all three cases, a simple linear program solver was used. The experiments show that the resulting networks maintain agreement with the original ANN and excellent performance on the evaluation tasks. The networks were also reduced in size with little loss in task performance.

Lindsey Kuper, Guy Katz, Justin Gottschlich et al.

The increasing use of deep neural networks for safety-critical applications, such as autonomous driving and flight control, raises concerns about their safety and reliability. Formal verification can address these concerns by guaranteeing that a deep learning system operates as intended, but the state of the art is limited to small systems. In this work-in-progress report we give an overview of our work on mitigating this difficulty, by pursuing two complementary directions: devising scalable verification techniques, and identifying design choices that result in deep learning systems that are more amenable to verification.

Zachary Sunberg, Mykel Kochenderfer

Online solvers for partially observable Markov decision processes have been applied to problems with large discrete state spaces, but continuous state, action, and observation spaces remain a challenge. This paper begins by investigating double progressive widening (DPW) as a solution to this challenge. However, we prove that this modification alone is not sufficient because the belief representations in the search tree collapse to a single particle causing the algorithm to converge to a policy that is suboptimal regardless of the computation time. This paper proposes and evaluates two new algorithms, POMCPOW and PFT-DPW, that overcome this deficiency by using weighted particle filtering. Simulation results show that these modifications allow the algorithms to be successful where previous approaches fail.

Patrick Slade, Preston Culbertson, Zachary Sunberg et al.

Robots performing manipulation tasks must operate under uncertainty about both their pose and the dynamics of the system. In order to remain robust to modeling error and shifts in payload dynamics, agents must simultaneously perform estimation and control tasks. However, the optimal estimation actions are often not the optimal actions for accomplishing the control tasks, and thus agents trade between exploration and exploitation. This work frames the problem as a Bayes-adaptive Markov decision process and solves it online using Monte Carlo tree search and an extended Kalman filter to handle Gaussian process noise and parameter uncertainty in a continuous space. MCTS selects control actions to reduce model uncertainty and reach the goal state nearly optimally. Certainty equivalent model predictive control is used as a benchmark to compare performance in simulations with varying process noise and parameter uncertainty.

Guy Katz, Clark Barrett, David Dill et al.

Deep neural networks have emerged as a widely used and effective means for tackling complex, real-world problems. However, a major obstacle in applying them to safety-critical systems is the great difficulty in providing formal guarantees about their behavior. We present a novel, scalable, and efficient technique for verifying properties of deep neural networks (or providing counter-examples). The technique is based on the simplex method, extended to handle the non-convex Rectified Linear Unit (ReLU) activation function, which is a crucial ingredient in many modern neural networks. The verification procedure tackles neural networks as a whole, without making any simplifying assumptions. We evaluated our technique on a prototype deep neural network implementation of the next-generation airborne collision avoidance system for unmanned aircraft (ACAS Xu). Results show that our technique can successfully prove properties of networks that are an order of magnitude larger than the largest networks verified using existing methods.

Zachary Sunberg, Christopher Ho, Mykel Kochenderfer

Safe interaction with human drivers is one of the primary challenges for autonomous vehicles. In order to plan driving maneuvers effectively, the vehicle's control system must infer and predict how humans will behave based on their latent internal state (e.g., intentions and aggressiveness). This research uses a simple model for human behavior with unknown parameters that make up the internal states of the traffic participants and presents a method for quantifying the value of estimating these states and planning with their uncertainty explicitly modeled. An upper performance bound is established by an omniscient Monte Carlo Tree Search (MCTS) planner that has perfect knowledge of the internal states. A baseline lower bound is established by planning with MCTS assuming that all drivers have the same internal state. MCTS variants are then used to solve a partially observable Markov decision process (POMDP) that models the internal state uncertainty to determine whether inferring the internal state offers an advantage over the baseline. Applying this method to a freeway lane changing scenario reveals that there is a significant performance gap between the upper bound and baseline. POMDP planning techniques come close to closing this gap, especially when important hidden model parameters are correlated with measurable parameters.

Tim Allan Wheeler, Martin Holder, Hermann Winner et al.

Accurate simulation and validation of advanced driver assistance systems requires accurate sensor models. Modeling automotive radar is complicated by effects such as multipath reflections, interference, reflective surfaces, discrete cells, and attenuation. Detailed radar simulations based on physical principles exist but are computationally intractable for realistic automotive scenes. This paper describes a methodology for the construction of stochastic automotive radar models based on deep learning with adversarial loss connected to real-world data. The resulting model exhibits fundamental radar effects while remaining real-time capable.

Alex Kuefler, Jeremy Morton, Tim Wheeler et al.

The ability to accurately predict and simulate human driving behavior is critical for the development of intelligent transportation systems. Traditional modeling methods have employed simple parametric models and behavioral cloning. This paper adopts a method for overcoming the problem of cascading errors inherent in prior approaches, resulting in realistic behavior that is robust to trajectory perturbations. We extend Generative Adversarial Imitation Learning to the training of recurrent policies, and we demonstrate that our model outperforms rule-based controllers and maximum likelihood models in realistic highway simulations. Our model both reproduces emergent behavior of human drivers, such as lane change rate, while maintaining realistic control over long time horizons.