Search

Learning-based approaches have recently shown notable success in legged locomotion. However, these approaches often lack accountability, necessitating empirical tests to determine their effectiveness. In this work, we are interested in designing a learning-based locomotion controller whose stability can be examined and guaranteed. This can be achieved by verifying regions of attraction (RoAs) of legged robots to their stable walking gaits. This is a non-trivial problem for legged robots due to their hybrid dynamics. Although previous work has shown the utility of Hamilton-Jacobi (HJ) reachability to solve this problem, its practicality was limited by its poor scalability. The core contribution of our work is the employment of a deep learning-based HJ reachability solution to the hybrid legged robot dynamics, which overcomes the previous work's limitation. With the learned reachability solution, first, we can estimate a library of RoAs for various gaits. Second, we can design a one-step predictive controller that effectively stabilizes to an individual gait within the verified RoA. Finally, we can devise a strategy that switches gaits, in response to external perturbations, whose feasibility is guided by the RoA analysis. We demonstrate our method in a two-link walker simulation, whose mathematical model is well established. Our method achieves improved stability than previous model-based methods, while ensuring transparency that was not present in the existing learning-based approaches., Comment: The first two authors contributed equally. This work is supported in part by the NSF Grant CMMI-1944722, the NSF CAREER Program under award 2240163, the NASA ULI on Safe Aviation Autonomy, and the DARPA Assured Autonomy and Assured Neuro Symbolic Learning and Reasoning (ANSR) programs. The work of Jason J. Choi received the support of a fellowship from Kwanjeong Educational Foundation, Korea

Perception-based navigation systems are useful for unmanned ground vehicle (UGV) navigation in complex terrains, where traditional depth-based navigation schemes are insufficient. However, these data-driven methods are highly dependent on their training data and can fail in surprising and dramatic ways with little warning. To ensure the safety of the vehicle and the surrounding environment, it is imperative that the navigation system is able to recognize the predictive uncertainty of the perception model and respond safely and effectively in the face of uncertainty. In an effort to enable safe navigation under perception uncertainty, we develop a probabilistic and reconstruction-based competency estimation (PaRCE) method to estimate the model's level of familiarity with an input image as a whole and with specific regions in the image. We find that the overall competency score can correctly predict correctly classified, misclassified, and out-of-distribution (OOD) samples. We also confirm that the regional competency maps can accurately distinguish between familiar and unfamiliar regions across images. We then use this competency information to develop a planning and control scheme that enables effective navigation while maintaining a low probability of error. We find that the competency-aware scheme greatly reduces the number of collisions with unfamiliar obstacles, compared to a baseline controller with no competency awareness. Furthermore, the regional competency information is very valuable in enabling efficient navigation.

We propose a new reachability learning framework for high-dimensional nonlinear systems, focusing on reach-avoid problems. These problems require computing the reach-avoid set, which ensures that all its elements can safely reach a target set despite any disturbance within pre-specified bounds. Our framework has two main parts: offline learning of a newly designed reach-avoid value function and post-learning certification. Compared to prior works, our new value function is Lipschitz continuous and its associated Bellman operator is a contraction mapping, both of which improve the learning performance. To ensure deterministic guarantees of our learned reach-avoid set, we introduce two efficient post-learning certification methods. Both methods can be used online for real-time local certification or offline for comprehensive certification. We validate our framework in a 12-dimensional crazyflie drone racing hardware experiment and a simulated 10-dimensional highway takeover example., Comment: Submitted, under review

Power grids are seeing more devices connected at the load level in the form of power electronics: e.g., data centers, electric vehicle chargers, and battery storage facilities. Therefore it is necessary to perform power system analyses with load models that capture these loads' behavior, which has historically not been done. To this end, we propose ZIP-E loads, a composite load model that has a ZIP load with a dynamic power electronic, or E, load model. We perform small signal and transient analysis of the IEEE WSCC 9 Bus test case with ZIP and ZIP-E load models. For small signals, we conclude that ZIP loads destabalize networks significantly faster than corresponding ZIP-E loads. In stable cases, transient results showed significantly larger oscillations for ZIP loads. Further, we find that a higher network loading condition is correlated with a higher sensitivity to load model choice. These results suggests that the constant power portion of the ZIP load has a large destabilizing effect and can generally overestimate instability, and that attention should be drawn to load model choice if operating near a stability boundary.

While deep neural network (DNN)-based perception models are useful for many applications, these models are black boxes and their outputs are not yet well understood. To confidently enable a real-world, decision-making system to utilize such a perception model without human intervention, we must enable the system to reason about the perception model's level of competency and respond appropriately when the model is incompetent. In order for the system to make an intelligent decision about the appropriate action when the model is incompetent, it would be useful for the system to understand why the model is incompetent. We explore five novel methods for identifying regions in the input image contributing to low model competency, which we refer to as image cropping, segment masking, pixel perturbation, competency gradients, and reconstruction loss. We assess the ability of these five methods to identify unfamiliar objects, recognize regions associated with unseen classes, and identify unexplored areas in an environment. We find that the competency gradients and reconstruction loss methods show great promise in identifying regions associated with low model competency, particularly when aspects of the image that are unfamiliar to the perception model are causing this reduction in competency. Both of these methods boast low computation times and high levels of accuracy in detecting image regions that are unfamiliar to the model, allowing them to provide potential utility in decision-making pipelines. The code for reproducing our methods and results is available on GitHub: https://github.com/sarapohland/explainable-competency.

Reinforcement learning (RL) methods for social robot navigation show great success navigating robots through large crowds of people, but the performance of these learning-based methods tends to degrade in particularly challenging or unfamiliar situations due to the models' dependency on representative training data. To ensure human safety and comfort, it is critical that these algorithms handle uncommon cases appropriately, but the low frequency and wide diversity of such situations present a significant challenge for these data-driven methods. To overcome this challenge, we propose modifications to the learning process that encourage these RL policies to maintain additional caution in unfamiliar situations. Specifically, we improve the Socially Attentive Reinforcement Learning (SARL) policy by (1) modifying the training process to systematically introduce deviations into a pedestrian model, (2) updating the value network to estimate and utilize pedestrian-unpredictability features, and (3) implementing a reward function to learn an effective response to pedestrian unpredictability. Compared to the original SARL policy, our modified policy maintains similar navigation times and path lengths, while reducing the number of collisions by 82% and reducing the proportion of time spent in the pedestrians' personal space by up to 19 percentage points for the most difficult cases. We also describe how to apply these modifications to other RL policies and demonstrate that some key high-level behaviors of our approach transfer to a physical robot.

Game-playing agents like AlphaGo have achieved superhuman performance through self-play, which is theoretically guaranteed to yield optimal policies in competitive games. However, most language tasks are partially or fully cooperative, so it is an open question whether techniques like self-play can effectively be used to improve language models. We empirically investigate this question in a negotiation game setting known as Deal or No Deal (DoND). Crucially, the objective in DoND can be modified to produce a fully cooperative game, a strictly competitive one, or anything in between. We finetune language models in self-play over multiple rounds of filtered behavior cloning in DoND for each of these objectives. Contrary to expectations, we find that language model self-play leads to significant performance gains in both cooperation and competition with humans, suggesting that self-play and related techniques have promise despite a lack of theoretical guarantees.

We show that domain-general automatic evaluators can significantly improve the performance of agents for web navigation and device control. We experiment with multiple evaluation models that trade off between inference cost, modularity of design, and accuracy. We validate the performance of these models in several popular benchmarks for digital agents, finding between 74.4 and 92.9% agreement with oracle evaluation metrics. Finally, we use these evaluators to improve the performance of existing agents via fine-tuning and inference-time guidance. Without any additional supervision, we improve state-of-the-art performance by 29% on the popular benchmark WebArena, and achieve a 75% relative improvement in a challenging domain transfer scenario., Comment: Code at https://github.com/Berkeley-NLP/Agent-Eval-Refine

Large language models are known to hallucinate when faced with unfamiliar queries, but the underlying mechanism that govern how models hallucinate are not yet fully understood. In this work, we find that unfamiliar examples in the models' finetuning data -- those that introduce concepts beyond the base model's scope of knowledge -- are crucial in shaping these errors. In particular, we find that an LLM's hallucinated predictions tend to mirror the responses associated with its unfamiliar finetuning examples. This suggests that by modifying how unfamiliar finetuning examples are supervised, we can influence a model's responses to unfamiliar queries (e.g., say ``I don't know''). We empirically validate this observation in a series of controlled experiments involving SFT, RL, and reward model finetuning on TriviaQA and MMLU. Our work further investigates RL finetuning strategies for improving the factuality of long-form model generations. We find that, while hallucinations from the reward model can significantly undermine the effectiveness of RL factuality finetuning, strategically controlling how reward models hallucinate can minimize these negative effects. Leveraging our previous observations on controlling hallucinations, we propose an approach for learning more reliable reward models, and show that they improve the efficacy of RL factuality finetuning in long-form biography and book/movie plot generation tasks.

Autonomous agents should be able to coordinate with other agents without knowing their intents ahead of time. While prior work has studied how agents can gather information about the intent of others, in this work we study the inverse problem: how agents can demonstrate their intent to others, within the framework of general-sum dynamic games. We first present a model of this intent demonstration problem and then propose an algorithm that enables an agent to trade off their task performance and intent demonstration to improve the overall system's performance. To scale to continuous states and action spaces as well as to nonlinear dynamics and costs, our algorithm leverages linear-quadratic approximations with an efficient intent teaching guarantee. Our empirical results show that intent demonstration accelerates other agents' learning and enables the demonstrating agent to balance task performance with intent expression., Comment: Under review by L4DC 2024

Learning-based approaches are emerging as an effective approach for safety filters for black-box dynamical systems. Existing methods have relied on certificate functions like Control Barrier Functions (CBFs) and Hamilton-Jacobi (HJ) reachability value functions. The primary motivation for our work is the recognition that ultimately, enforcing the safety constraint as a control input constraint at each state is what matters. By focusing on this constraint, we can eliminate dependence on any specific certificate function-based design. To achieve this, we define a discriminating hyperplane that shapes the half-space constraint on control input at each state, serving as a sufficient condition for safety. This concept not only generalizes over traditional safety methods but also simplifies safety filter design by eliminating dependence on specific certificate functions. We present two strategies to learn the discriminating hyperplane: (a) a supervised learning approach, using pre-verified control invariant sets for labeling, and (b) a reinforcement learning (RL) approach, which does not require such labels. The main advantage of our method, unlike conventional safe RL approaches, is the separation of performance and safety. This offers a reusable safety filter for learning new tasks, avoiding the need to retrain from scratch. As such, we believe that the new notion of the discriminating hyperplane offers a more generalizable direction towards designing safety filters, encompassing and extending existing certificate-function-based or safe RL methodologies., Comment: * Indicate co-first authors. This is an extended version of the paper presented at L4DC 2024

Solving feedback Stackelberg games with nonlinear dynamics and coupled constraints, a common scenario in practice, presents significant challenges. This work introduces an efficient method for computing approximate local feedback Stackelberg equilibria in multi-player general-sum dynamic games, with continuous state and action spaces. Different from existing (approximate) dynamic programming solutions that are primarily designed for unconstrained problems, our approach involves reformulating a feedback Stackelberg dynamic game into a sequence of nested optimization problems, enabling the derivation of Karush-Kuhn-Tucker (KKT) conditions and the establishment of a second-order sufficient condition for local feedback Stackelberg equilibria. We propose a Newton-style primal-dual interior point method for solving constrained linear quadratic (LQ) feedback Stackelberg games, offering provable convergence guarantees. Our method is further extended to compute local feedback Stackelberg equilibria for more general nonlinear games by iteratively approximating them using LQ games, ensuring that their KKT conditions are locally aligned with those of the original nonlinear games. We prove the exponential convergence of our algorithm in constrained nonlinear games. In a feedback Stackelberg game with nonlinear dynamics and (nonconvex) coupled costs and constraints, our experimental results reveal the algorithm's ability to handle infeasible initial conditions and achieve exponential convergence towards an approximate local feedback Stackelberg equilibrium., Comment: This manuscript is currently under review by SIAM Journal on Optimization

Adversarial attacks on learning-based multi-modal trajectory predictors have already been demonstrated. However, there are still open questions about the effects of perturbations on inputs other than state histories, and how these attacks impact downstream planning and control. In this paper, we conduct a sensitivity analysis on two trajectory prediction models, Trajectron++ and AgentFormer. The analysis reveals that between all inputs, almost all of the perturbation sensitivities for both models lie only within the most recent position and velocity states. We additionally demonstrate that, despite dominant sensitivity on state history perturbations, an undetectable image map perturbation made with the Fast Gradient Sign Method can induce large prediction error increases in both models, revealing that these trajectory predictors are, in fact, susceptible to image-based attacks. Using an optimization-based planner and example perturbations crafted from sensitivity results, we show how these attacks can cause a vehicle to come to a sudden stop from moderate driving speeds., Comment: 10 pages, 5 figures, 1 tables

As the use of autonomous robots expands in tasks that are complex and challenging to model, the demand for robust data-driven control methods that can certify safety and stability in uncertain conditions is increasing. However, the practical implementation of these methods often faces scalability issues due to the growing amount of data points with system complexity, and a significant reliance on high-quality training data. In response to these challenges, this study presents a scalable data-driven controller that efficiently identifies and infers from the most informative data points for implementing data-driven safety filters. Our approach is grounded in the integration of a model-based certificate function-based method and Gaussian Process (GP) regression, reinforced by a novel online data selection algorithm that reduces time complexity from quadratic to linear relative to dataset size. Empirical evidence, gathered from successful real-world cart-pole swing-up experiments and simulated locomotion of a five-link bipedal robot, demonstrates the efficacy of our approach. Our findings reveal that our efficient online data selection algorithm, which strategically selects key data points, enhances the practicality and efficiency of data-driven certifying filters in complex robotic systems, significantly mitigating scalability concerns inherent in nonparametric learning-based control methods., Comment: The first three authors contributed equally to the work. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible

Lightcurves of many classical novae deviate from the canonical "fast rise - smooth decline" pattern and display complex variability behavior. We present the first TESS-space-photometry-based investigation of this phenomenon. We use TESS Sector 41 full-frame images to extract a lightcurve of the slow Galactic nova V606 Vul that erupted nine days prior to the start of the TESS observations. The lightcurve covers the first of two major peaks of V606 Vul that was reached 19 days after the start of the eruption. The nova reached its brightest visual magnitude V=9.9 in its second peak 64 days after the eruption onset, following the completion of Sector 41 observations. To increase the confidence level of the extracted lightcurve, we performed the analysis using four different codes implementing the aperture photometry (Lightkurve, VaST) and image subtraction (TESSreduce, tequila_shots) and find good agreement between them. We performed ground-based photometric and spectroscopic monitoring to complement the TESS data. The TESS lightcurve reveals two features: periodic variations (0.12771d, 0.01mag average peak-to-peak amplitude) that disappeared when the source was within 1mag of peak optical brightness and a series of isolated mini-flares (with peak-to-peak amplitudes of up to 0.5mag) appearing at seemingly random times. We interpret the periodic variations as the result of azimuthal asymmetry of the photosphere engulfing the nova-hosting binary that was distorted by and rotating with the binary. Whereas we use spectra to associate the two major peaks in the nova lightcurve with distinct episodes of mass ejection, the origin of mini-flares remains elusive., Comment: 31 pages, 10 figures, submitted to ApJ; comments welcome

Control invariant sets are crucial for various methods that aim to design safe control policies for systems whose state constraints must be satisfied over an indefinite time horizon. In this article, we explore the connections among reachability, control invariance, and Control Barrier Functions (CBFs). Unlike prior formulations based on backward reachability concepts, by examining a forward reachability problem, we are able to establish a strong link between these three concepts. First, our findings show that the inevitable Forward Reachable Tube (FRT), which is the set of states such that every trajectory reaching the FRT must have passed through a given initial set of states, is precisely this initial set of states itself if it is a robust control invariant set with a differentiable boundary. We highlight that this statement may not hold if the boundary is not differentiable. Next, we formulate a differential game between the control and disturbance, where the inevitable FRT is characterized by the zero-superlevel set of the value function. By incorporating a discount factor in the cost function of the game, the barrier constraint of the CBF naturally arises as the constraint that is imposed on the optimal control policy. Combining these results, the value function of our FRT formulation serves as a CBF-like function, and conversely, any valid CBF is also a forward reachability value function inside the control invariant set, thereby revealing the inverse optimality of the CBF. This strong link we establish between the reachability problem and the barrier constraint, while guaranteeing the continuity of the value function, is not achievable by previous backward reachability-based formulations. As such, our work fills a crucial gap in the existing literature that is vital for constructing valid CBFs to ensure safety., Comment: The first two authors contributed equally to this work

Conventional wisdom suggests that neural network predictions tend to be unpredictable and overconfident when faced with out-of-distribution (OOD) inputs. Our work reassesses this assumption for neural networks with high-dimensional inputs. Rather than extrapolating in arbitrary ways, we observe that neural network predictions often tend towards a constant value as input data becomes increasingly OOD. Moreover, we find that this value often closely approximates the optimal constant solution (OCS), i.e., the prediction that minimizes the average loss over the training data without observing the input. We present results showing this phenomenon across 8 datasets with different distributional shifts (including CIFAR10-C and ImageNet-R, S), different loss functions (cross entropy, MSE, and Gaussian NLL), and different architectures (CNNs and transformers). Furthermore, we present an explanation for this behavior, which we first validate empirically and then study theoretically in a simplified setting involving deep homogeneous networks with ReLU activations. Finally, we show how one can leverage our insights in practice to enable risk-sensitive decision-making in the presence of OOD inputs.

In order for robots to safely navigate in unseen scenarios using learning-based methods, it is important to accurately detect out-of-training-distribution (OoD) situations online. Recently, Gaussian process state-space models (GPSSMs) have proven useful to discriminate unexpected observations by comparing them against probabilistic predictions. However, the capability for the model to correctly distinguish between in- and out-of-training distribution observations hinges on the accuracy of these predictions, primarily affected by the class of functions the GPSSM kernel can represent. In this paper, we propose (i) a novel approach to embed existing domain knowledge in the kernel and (ii) an OoD online runtime monitor, based on receding-horizon predictions. Domain knowledge is provided in the form of a dataset, collected either in simulation or by using a nominal model. Numerical results show that the informed kernel yields better regression quality with smaller datasets, as compared to standard kernel choices. We demonstrate the effectiveness of the OoD monitor on a real quadruped navigating an indoor setting, which reliably classifies previously unseen terrains., Comment: 7 pages, 4 figures

Object-centric representations enable autonomous driving algorithms to reason about interactions between many independent agents and scene features. Traditionally these representations have been obtained via supervised learning, but this decouples perception from the downstream driving task and could harm generalization. In this work we adapt a self-supervised object-centric vision model to perform object decomposition using only RGB video and the pose of the vehicle as inputs. We demonstrate that our method obtains promising results on the Waymo Open perception dataset. While object mask quality lags behind supervised methods or alternatives that use more privileged information, we find that our model is capable of learning a representation that fuses multiple camera viewpoints over time and successfully tracks many vehicles and pedestrians in the dataset. Code for our model is available at https://github.com/wayveai/SOCS., Comment: Presented at the CVPR 2023 Vision-Centric Autonomous Driving workshop

Autonomous multi-agent systems are increasingly being deployed in environments where winds and ocean currents have a significant influence. Recent work has developed control policies for single agents that leverage flows to achieve their objectives in dynamic environments. However, in multi-agent systems, these flows can cause agents to collide or drift apart and lose direct inter-agent communications, especially when agents have low propulsion capabilities. To address these challenges, we propose a hierarchical multi-agent control approach that allows arbitrary single-agent performance policies that are unaware of other agents to be used in multi-agent systems while ensuring safe operation. We first develop a safety controller using potential functions, solely dedicated to avoiding collisions and maintaining inter-agent communication. Next, we design a low-interference safe interaction (LISIC) policy that trades off the performance policy and the safety control to ensure safe and performant operation. Specifically, when the agents are at an appropriate distance, LISIC prioritizes the performance policy while smoothly increasing the safety controller when necessary. We prove that under mild assumptions on the flows experienced by the agents, our approach can guarantee safety. Additionally, we demonstrate the effectiveness of our method in realistic settings through an extensive empirical analysis with simulations of fleets of underactuated autonomous surface vehicles operating in dynamic ocean currents where these assumptions do not always hold., Comment: 8 pages, Published at European Control Conference 2024 (ECC 2024) Nicolas Hoischen and Marius Wiggert contributed equally to this work

Low-propulsion vessels can take advantage of powerful ocean currents to navigate towards a destination. Recent results demonstrated that vessels can reach their destination with high probability despite forecast errors. However, these results do not consider the critical aspect of safety of such vessels: because of their low propulsion which is much smaller than the magnitude of currents, they might end up in currents that inevitably push them into unsafe areas such as shallow areas, garbage patches, and shipping lanes. In this work, we first investigate the risk of stranding for free-floating vessels in the Northeast Pacific. We find that at least 5.04% would strand within 90 days. Next, we encode the unsafe sets as hard constraints into Hamilton-Jacobi Multi-Time Reachability (HJ-MTR) to synthesize a feedback policy that is equivalent to re-planning at each time step at low computational cost. While applying this policy closed-loop guarantees safe operation when the currents are known, in realistic situations only imperfect forecasts are available. We demonstrate the safety of our approach in such realistic situations empirically with large-scale simulations of a vessel navigating in high-risk regions in the Northeast Pacific. We find that applying our policy closed-loop with daily re-planning on new forecasts can ensure safety with high probability even under forecast errors that exceed the maximal propulsion. Our method significantly improves safety over the baselines and still achieves a timely arrival of the vessel at the destination., Comment: 6 pages, 3 figures, submitted to 2023 IEEE 62th Annual Conference on Decision and Control (CDC) Andreas Doering and Marius Wiggert contributed equally to this work

Seaweed biomass offers significant potential for climate mitigation, but large-scale, autonomous open-ocean farms are required to fully exploit it. Such farms typically have low propulsion and are heavily influenced by ocean currents. We want to design a controller that maximizes seaweed growth over months by taking advantage of the non-linear time-varying ocean currents for reaching high-growth regions. The complex dynamics and underactuation make this challenging even when the currents are known. This is even harder when only short-term imperfect forecasts with increasing uncertainty are available. We propose a dynamic programming-based method to efficiently solve for the optimal growth value function when true currents are known. We additionally present three extensions when as in reality only forecasts are known: (1) our methods resulting value function can be used as feedback policy to obtain the growth-optimal control for all states and times, allowing closed-loop control equivalent to re-planning at every time step hence mitigating forecast errors, (2) a feedback policy for long-term optimal growth beyond forecast horizons using seasonal average current data as terminal reward, and (3) a discounted finite-time Dynamic Programming (DP) formulation to account for increasing ocean current estimate uncertainty. We evaluate our approach through 30-day simulations of floating seaweed farms in realistic Pacific Ocean current scenarios. Our method demonstrates an achievement of 95.8% of the best possible growth using only 5-day forecasts. This confirms the feasibility of using low-power propulsion and optimal control for enhanced seaweed growth on floating farms under real-world conditions., Comment: 8 pages, submitted to 2023 IEEE 62th Annual Conference on Decision and Control (CDC) Matthias Killer and Marius Wiggert contributed equally to this work

We describe a class of tasks called decision-oriented dialogues, in which AI assistants such as large language models (LMs) must collaborate with one or more humans via natural language to help them make complex decisions. We formalize three domains in which users face everyday decisions: (1) choosing an assignment of reviewers to conference papers, (2) planning a multi-step itinerary in a city, and (3) negotiating travel plans for a group of friends. In each of these settings, AI assistants and users have disparate abilities that they must combine to arrive at the best decision: assistants can access and process large amounts of information, while users have preferences and constraints external to the system. For each task, we build a dialogue environment where agents receive a reward based on the quality of the final decision they reach. We evaluate LMs in self-play and in collaboration with humans and find that they fall short compared to human assistants, achieving much lower rewards despite engaging in longer dialogues. We highlight a number of challenges models face in decision-oriented dialogues, ranging from goal-directed behavior to reasoning and optimization, and release our environments as a testbed for future work., Comment: TACL 2024, pre-MIT Press publication version

We introduce Ghostbuster, a state-of-the-art system for detecting AI-generated text. Our method works by passing documents through a series of weaker language models, running a structured search over possible combinations of their features, and then training a classifier on the selected features to predict whether documents are AI-generated. Crucially, Ghostbuster does not require access to token probabilities from the target model, making it useful for detecting text generated by black-box models or unknown model versions. In conjunction with our model, we release three new datasets of human- and AI-generated text as detection benchmarks in the domains of student essays, creative writing, and news articles. We compare Ghostbuster to a variety of existing detectors, including DetectGPT and GPTZero, as well as a new RoBERTa baseline. Ghostbuster achieves 99.0 F1 when evaluated across domains, which is 5.9 F1 higher than the best preexisting model. It also outperforms all previous approaches in generalization across writing domains (+7.5 F1), prompting strategies (+2.1 F1), and language models (+4.4 F1). We also analyze the robustness of our system to a variety of perturbations and paraphrasing attacks and evaluate its performance on documents written by non-native English speakers., Comment: NAACL 2024

The uniform information density (UID) hypothesis states that humans tend to distribute information roughly evenly across an utterance or discourse. Early evidence in support of the UID hypothesis came from Genzel & Charniak (2002), which proposed an entropy rate constancy principle based on the probability of English text under n-gram language models. We re-evaluate the claims of Genzel & Charniak (2002) with neural language models, failing to find clear evidence in support of entropy rate constancy. We conduct a range of experiments across datasets, model sizes, and languages and discuss implications for the uniform information density hypothesis and linguistic theories of efficient communication more broadly., Comment: Findings of EMNLP 2023

Decision-making in multi-player games can be extremely challenging, particularly under uncertainty. In this work, we propose a new sample-based approximation to a class of stochastic, general-sum, pure Nash games, where each player has an expected-value objective and a set of chance constraints. This new approximation scheme inherits the accuracy of objective approximation from the established sample average approximation (SAA) method and enjoys a feasibility guarantee derived from the scenario optimization literature. We characterize the sample complexity of this new game-theoretic approximation scheme, and observe that high accuracy usually requires a large number of samples, which results in a large number of sampled constraints. To accommodate this, we decompose the approximated game into a set of smaller games with few constraints for each sampled scenario, and propose a decentralized, consensus-based ADMM algorithm to efficiently compute a generalized Nash equilibrium (GNE) of the approximated game. We prove the convergence of our algorithm to a GNE and empirically demonstrate superior performance relative to a recent baseline algorithm based on ADMM and interior point method.

We investigate methods to provide safety assurances for autonomous agents that incorporate predictions of other, uncontrolled agents' behavior into their own trajectory planning. Given a learning-based forecasting model that predicts agents' trajectories, we introduce a method for providing probabilistic assurances on the model's prediction error with calibrated confidence intervals. Through quantile regression, conformal prediction, and reachability analysis, our method generates probabilistically safe and dynamically feasible prediction sets. We showcase their utility in certifying the safety of planning algorithms, both in simulations using actual autonomous driving data and in an experiment with Boeing vehicles.

Robots deployed to the real world must be able to interact with other agents in their environment. Dynamic game theory provides a powerful mathematical framework for modeling scenarios in which agents have individual objectives and interactions evolve over time. However, a key limitation of such techniques is that they require a-priori knowledge of all players' objectives. In this work, we address this issue by proposing a novel method for learning players' objectives in continuous dynamic games from noise-corrupted, partial state observations. Our approach learns objectives by coupling the estimation of unknown cost parameters of each player with inference of unobserved states and inputs through Nash equilibrium constraints. By coupling past state estimates with future state predictions, our approach is amenable to simultaneous online learning and prediction in receding horizon fashion. We demonstrate our method in several simulated traffic scenarios in which we recover players' preferences for, e.g., desired travel speed and collision-avoidance behavior. Results show that our method reliably estimates game-theoretic models from noise-corrupted data that closely matches ground-truth objectives, consistently outperforming state-of-the-art approaches., Comment: arXiv admin note: text overlap with arXiv:2106.03611

In multi-agent dynamic games, the Nash equilibrium state trajectory of each agent is determined by its cost function and the information pattern of the game. However, the cost and trajectory of each agent may be unavailable to the other agents. Prior work on using partial observations to infer the costs in dynamic games assumes an open-loop information pattern. In this work, we demonstrate that the feedback Nash equilibrium concept is more expressive and encodes more complex behavior. It is desirable to develop specific tools for inferring players' objectives in feedback games. Therefore, we consider the dynamic game cost inference problem under the feedback information pattern, using only partial state observations and incomplete trajectory data. To this end, we first propose an inverse feedback game loss function, whose minimizer yields a feedback Nash equilibrium state trajectory closest to the observation data. We characterize the landscape and differentiability of the loss function. Given the difficulty of obtaining the exact gradient, our main contribution is an efficient gradient approximator, which enables a novel inverse feedback game solver that minimizes the loss using first-order optimization. In thorough empirical evaluations, we demonstrate that our algorithm converges reliably and has better robustness and generalization performance than the open-loop baseline method when the observation data reflects a group of players acting in a feedback Nash game., Comment: Accepted by AAMAS 2023. This is a preprint version

We study representation learning for efficient imitation learning over linear systems. In particular, we consider a setting where learning is split into two phases: (a) a pre-training step where a shared $k$-dimensional representation is learned from $H$ source policies, and (b) a target policy fine-tuning step where the learned representation is used to parameterize the policy class. We find that the imitation gap over trajectories generated by the learned target policy is bounded by $\tilde{O}\left( \frac{k n_x}{HN_{\mathrm{shared}}} + \frac{k n_u}{N_{\mathrm{target}}}\right)$, where $n_x > k$ is the state dimension, $n_u$ is the input dimension, $N_{\mathrm{shared}}$ denotes the total amount of data collected for each policy during representation learning, and $N_{\mathrm{target}}$ is the amount of target task data. This result formalizes the intuition that aggregating data across related tasks to learn a representation can significantly improve the sample efficiency of learning a target task. The trends suggested by this bound are corroborated in simulation., Comment: Appeared in L4DC 2023. V3: corrected typo in assumptions