Robot learning from humans has been proposed and researched for several decades as a means to enable robots to learn new skills or adapt existing ones to new situations. Recent advances in AI, including learning approaches like reinforcement learning and architectures like transformers and foundation models, combined with access to massive datasets, have created attractive opportunities to apply those data-hungry techniques to this problem. We argue that the focus on massive amounts of pre-collected data, and the resulting learning paradigm, where humans demonstrate and robots learn in isolation, is overshadowing a specialized area of work we term Human-Interactive Robot Learning (HIRL). This paradigm, wherein robots and humans interact during the learning process, is at the intersection of multiple fields (AI, robotics, human-computer interaction, design and others) and holds unique promise. Using HIRL, robots can achieve greater sample efficiency (as humans can provide task knowledge through interaction), align with human preferences (as humans can guide the robot behavior toward their expectations), and explore more meaningfully and safely (as humans can utilize domain knowledge to guide learning and prevent catastrophic failures). This can result in robotic systems that can more quickly and easily adapt to new tasks in human environments. The objective of this article is to provide a broad and consistent overview of HIRL research and to guide researchers toward understanding the scope of HIRL, and current open or underexplored challenges related to four themes-namely, human, robot learning, interaction, and broader context. The article includes concrete use cases to illustrate the interaction between these challenges and inspire further research according to broad recommendations and a call for action for the growing HIRL community.

Designing a reinforcement learning from human feedback (RLHF) algorithm to approximate a human's unobservable reward function requires assuming, implicitly or explicitly, a model of human preferences. A preference model that poorly describes how humans generate preferences risks learning a poor approximation of the human's reward function. In this paper, we conduct three human studies to asses whether one can influence the expression of real human preferences to more closely conform to a desired preference model. Importantly, our approach does not seek to alter the human's unobserved reward function. Rather, we change how humans use this reward function to generate preferences, such that they better match whatever preference model is assumed by a particular RLHF algorithm. We introduce three interventions: showing humans the quantities that underlie a preference model, which is normally unobservable information derived from the reward function; training people to follow a specific preference model; and modifying the preference elicitation question. All intervention types show significant effects, providing practical tools to improve preference data quality and the resultant alignment of the learned reward functions. Overall we establish a novel research direction in model alignment: designing interfaces and training interventions to increase human conformance with the modeling assumptions of the algorithm that will learn from their input.

Reinforcement learning agents are fundamentally limited by the quality of the reward functions they learn from, yet reward design is often overlooked under the assumption that a well-defined reward is readily available. However, in practice, designing rewards is difficult, and even when specified, evaluating their correctness is equally problematic: how do we know if a reward function is correctly specified? In our work, we address these challenges by focusing on reward alignment – assessing whether a reward function accurately encodes the preferences of a human stakeholder. As a concrete measure of reward alignment, we introduce the Trajectory Alignment Coefficient to quantify the similarity between a human stakeholder's ranking of trajectory distributions and those induced by a given reward function. We show that the Trajectory Alignment Coefficient exhibits desirable properties, such as not requiring access to a ground truth reward, invariance to potential-based reward shaping, and applicability to online RL. Additionally, in an 11 – person user study of RL practitioners, we found that access to the Trajectory Alignment Coefficient during reward selection led to statistically significant improvements. Compared to relying only on reward functions, our metric reduced cognitive workload by 1.5x, was preferred by 82

Consumer protection laws are designed to protect consumers from unethical business practices. In this position paper, I argue that these laws serve an emergent dual purpose: if appropriately enforced and strengthened, consumer protection laws can serve as an inalienable defense for AI safety. These laws are well established and can be enforced and strengthened to incentivize businesses to design and deploy safer AI systems. This position runs counter to two prevailing trends in AI policy. The first alternative position is that AI safety requires an entirely new set of focused laws to protect humanity's prosperity. Though I find these efforts valuable, I argue that such focused laws are both hard to write and easy to skirt. The second alternative position is that consumer protection is nothing more than red tape; I argue that existing laws dating back many decades have already reigned in some nefarious business practices related to the development and deployment of AI, and that the litigious society of the United States is well-positioned to use consumer protection laws to encourage new AI safety guardrails. This paper takes a tour of some existing consumer protection laws in the United States and their effects on the development and use of AI systems. This paper also calls to enforce and preserve these laws in a rapidly changing, de-regulatory political landscape.

We consider algorithms for learning reward functions from human preferences over pairs of trajectory segments, as used in reinforcement learning from human feedback (RLHF). Most recent work assumes that human preferences are generated based only upon the reward accrued within those segments, or their partial return. Recent work casts doubt on the validity of this assumption, proposing an alternative preference model based upon regret. We investigate the consequences of assuming preferences are based upon partial return when they actually arise from regret. We argue that the learned function is an approximation of the optimal advantage function, not a reward function. We find that if a specific pitfall is addressed, this incorrect assumption is not particularly harmful, resulting in a highly shaped reward function. Nonetheless, this incorrect usage of the approximation of the optimal advantage function is less desirable than the appropriate and simpler approach of greedy maximization of it. From the perspective of the regret preference model, we also provide a clearer interpretation of fine tuning contemporary large language models with RLHF. This paper overall provides insight regarding why learning under the partial return preference model tends to work so well in practice, despite it conforming poorly to how humans give preferences.

Generative models can serve as surrogates for some real data sources by creating synthetic training datasets, but in doing so they may transfer biases to downstream tasks. We focus on protecting quality and diversity when generating synthetic training datasets. We propose quality-diversity generative sampling (QDGS), a framework for sampling data uniformly across a user-defined measure space, despite the data coming from a biased generator. QDGS is a model-agnostic framework that uses prompt guidance to optimize a quality objective across measures of diversity for synthetically generated data, without fine-tuning the generative model. Using balanced synthetic datasets generated by QDGS, we first debias classifiers trained on color-biased shape datasets as a proof-of-concept. By applying QDGS to facial data synthesis, we prompt for desired semantic concepts, such as skin tone and age, to create an intersectional dataset with a combined blend of visual features. Leveraging this balanced data for training classifiers improves fairness while maintaining accuracy on facial recognition benchmarks. Code available at: https://github.com/Cylumn/qd-generative-sampling.

As robots are introduced to various domains and applications, Human-Robot Teaming (HRT) capabilities are essential. Such capabilities involve teaming with humans in\ on\ out-the-loop at different levels of abstraction, leveraging the complementing capabilities of humans and robots. This requires robotic systems with the ability to dynamically vary their level or degree of autonomy to collaborate with the human(s) efficiently and overcome various challenging circumstances. Variable Autonomy (VA) is an umbrella term encompassing such research, including but not limited to shared control and shared autonomy, mixed-initiative, adjustable autonomy, and sliding autonomy. This workshop is driven by the timely need to bring together VA-related research and practices that are often disconnected across different communities as the field is relatively young. The workshop's goal is to consolidate research in VA. To this end, and given the complexity and span of Human-Robot systems, this workshop will adopt a holistic trans-disciplinary approach aiming to a) identify and classify related common challenges and opportunities; b) identify the disciplines that need to come together to tackle the challenges; c) identify and define common terminology, approaches, methodologies, benchmarks, and metrics; d) define short- and long-term research goals for the community. To achieve these objectives, this workshop aims to bring together industry stakeholders, researchers from fields under the banner of VA, and specialists from other highly related fields such as human factors and psychology. The workshop will consist of a mix of invited talks, contributed papers, and an interactive discussion panel, toward a shared vision for VA.

Human-robot interaction is limited in large part by the challenge of writing correct specifications for robots. The research community wants alignment between humans' goals and robot behaviors, but this alignment is very hard to achieve. My research tackles this problem. I view alignment as the consequence of iterative design and ample testing, and I design methods in service of these processes. I first study how humans currently write reward functions, and I profile some of the typical errors they make when doing so. I then study how humans can inspect the behaviors robots learn from any given specification. A typical approach to this mandates unstructured or hand-designed test cases; I instead introduce a Bayesian inference method for finding behavior examples which cover information-rich test cases. Alongside finding these behavior examples, I study how these examples should be presented to the human through applying cognitive theories of human concept learning. For the remainder of my thesis, I am pursuing two open questions. My first question concerns how these components can be combined such that humans are able to iteratively design better behavioral specifications. My second question concerns how robots can better interpret humans' erroneous specifications and attempt to infer their true intent, in spite of the errors.

In reinforcement learning (RL), a reward function that aligns exactly with a task's true performance metric is often necessarily sparse. For example, a true task metric might encode a reward of 1 upon success and 0 otherwise. The sparsity of these true task metrics can make them hard to learn from, so in practice they are often replaced with alternative dense reward functions. These dense reward functions are typically designed by experts through an ad hoc process of trial and error. In this process, experts manually search for a reward function that improves performance with respect to the task metric while also enabling an RL algorithm to learn faster. This process raises the question of whether the same reward function is optimal for all algorithms, i.e., whether the reward function can be overfit to a particular algorithm. In this paper, we study the consequences of this wide yet unexamined practice of trial-and-error reward design. We first conduct computational experiments that confirm that reward functions can be overfit to learning algorithms and their hyperparameters. We then conduct a controlled observation study which emulates expert practitioners' typical experiences of reward design, in which we similarly find evidence of reward function overfitting. We also find that experts' typical approach to reward design---of adopting a myopic strategy and weighing the relative goodness of each state-action pair---leads to misdesign through invalid task specifications, since RL algorithms use cumulative reward rather than rewards for individual state-action pairs as an optimization target. Code, data: github.com/serenabooth/reward-design-perils

When interacting with a robot, humans form conceptual models (of varying quality) which capture how the robot behaves. These conceptual models form just from watching or interacting with the robot, with or without conscious thought. Some methods select and present robot behaviors to improve human conceptual model formation; nonetheless, these methods and HRI more broadly have not yet consulted cognitive theories of human concept learning. These validated theories offer concrete design guidance to support humans in developing conceptual models more quickly, accurately, and flexibly. Specifically, Analogical Transfer Theory and the Variation Theory of Learning have been successfully deployed in other fields, and offer new insights for the HRI community about the selection and presentation of robot behaviors. Using these theories, we review and contextualize 35 prior works in human-robot teaching and learning, and we assess how these works incorporate or omit the design implications of these theories. From this review, we identify new opportunities for algorithms and interfaces to help humans more easily learn conceptual models of robot behaviors, which in turn can help humans become more effective robot teachers and collaborators.

The utility of reinforcement learning is limited by the alignment of reward functions with the interests of human stakeholders. One promising method for alignment is to learn the reward function from human-generated preferences between pairs of trajectory segments, a type of reinforcement learning from human feedback (RLHF). These human preferences are typically assumed to be informed solely by partial return, the sum of rewards along each segment. We find this assumption to be flawed and propose modeling human preferences instead as informed by each segment's regret, a measure of a segment's deviation from optimal decision-making. Given infinitely many preferences generated according to regret, we prove that we can identify a reward function equivalent to the reward function that generated those preferences, and we prove that the previous partial return model lacks this identifiability property in multiple contexts. We empirically show that our proposed regret preference model outperforms the partial return preference model with finite training data in otherwise the same setting. Additionally, we find that our proposed regret preference model better predicts real human preferences and also learns reward functions from these preferences that lead to policies that are better human-aligned. Overall, this work establishes that the choice of preference model is impactful, and our proposed regret preference model provides an improvement upon a core assumption of recent research. We have open sourced our experimental code, the human preferences dataset we gathered, and our training and preference elicitation interfaces for gathering a such a dataset.

Feature attribution methods are popular in interpretable machine learning. These methods compute the attribution of each input feature to represent its importance, but there is no consensus on the definition of "attribution", leading to many competing methods with little systematic evaluation, complicated in particular by the lack of ground truth attribution. To address this, we propose a dataset modification procedure to induce such ground truth. Using this procedure, we evaluate three common methods: saliency maps, rationales, and attentions. We identify several deficiencies and add new perspectives to the growing body of evidence questioning the correctness and reliability of these methods applied on datasets in the wild. We further discuss possible avenues for remedy and recommend new attribution methods to be tested against ground truth before deployment. The code and appendix are available at https://yilunzhou.github.io/feature-attribution-evaluation/.

As robots are deployed in complex situations, engineers and end users must develop a holistic understanding of their behaviors, capabilities, and limitations. Some behaviors are directly optimized by the objective function. They often include success rate, completion time or energy consumption. Other behaviors -- e.g., collision avoidance, trajectory smoothness or motion legibility -- are typically emergent but equally important for safe and trustworthy deployment. Designing an objective which optimizes every aspect of robot behavior is hard. In this paper, we advocate for systematic analysis of a wide array of behaviors for holistic understanding of robot controllers and, to this end, propose a framework, RoCUS, which uses Bayesian posterior sampling to find situations where the robot controller exhibits user-specified behaviors, such as highly jerky motions. We use RoCUS to analyze three controller classes (deep learning models, rapidly exploring random trees and dynamical system formulations) on two domains (2D navigation and a 7 degree-of-freedom arm reaching), and uncover insights to further our understanding of these controllers and ultimately improve their designs.

Neural rationale models are popular for interpretable predictions of NLP tasks. In these, a selector extracts segments of the input text, called rationales, and passes these segments to a classifier for prediction. Since the rationale is the only information accessible to the classifier, it is plausibly defined as the explanation. Is such a characterization unconditionally correct? In this paper, we argue to the contrary, with both philosophical perspectives and empirical evidence suggesting that rationale models are, perhaps, less rational and interpretable than expected. We call for more rigorous evaluations of these models to ensure desired properties of interpretability are indeed achieved. The code for our experiments is at https://github.com/yimingz89/Neural-Rationale-Analysis.

Post-hoc explanation methods are gaining popularity for interpreting, understanding, and debugging neural networks. Most analyses using such methods explain decisions in response to inputs drawn from the test set. However, the test set may have few examples that trigger some model behaviors, such as high-confidence failures or ambiguous classifications. To address these challenges, we introduce a flexible model inspection framework: Bayes-TrEx. Given a data distribution, Bayes-TrEx finds in-distribution examples which trigger a specified prediction confidence. We demonstrate several use cases of Bayes-TrEx, including revealing highly confident (mis)classifications, visualizing class boundaries via ambiguous examples, understanding novel-class extrapolation behavior, and exposing neural network overconfidence. We use Bayes-TrEx to study classifiers trained on CLEVR, MNIST, and Fashion-MNIST, and we show that this framework enables more flexible holistic model analysis than just inspecting the test set. Code and supplemental material are available at https://github.com/serenabooth/Bayes-TrEx.

This paper describes IEEE P7001, a new draft standard on transparency of autonomous systems 1 . In the paper, we outline the development and structure of the draft standard. We present the rationale for transparency as a measurable, testable property. We outline five stakeholder groups: users, the general public and bystanders, safety certification agencies, incident/accident investigators and lawyers/expert witnesses, and explain the thinking behind the normative definitions of “levels” of transparency for each stakeholder group in P7001. The paper illustrates the application of P7001 through worked examples of both specification and assessment of fictional autonomous systems.

Practitioners from diverse occupations and backgrounds are increasingly using machine learning (ML) methods. Nonetheless, studies on ML Practitioners typically draw populations from Big Tech and academia, as researchers have easier access to these communities. Through this selection bias, past research often excludes the broader, lesser-resourced ML community -- for example, practitioners working at startups, at non-tech companies, and in the public sector. These practitioners share many of the same ML development difficulties and ethical conundrums as their Big Tech counterparts; however, their experiences are subject to additional under-studied challenges stemming from deploying ML with limited resources, increased existential risk, and absent access to in-house research teams. We contribute a qualitative analysis of 17 interviews with stakeholders from organizations which are less represented in prior studies. We uncover a number of tensions which are introduced or exacerbated by these organizations' resource constraints -- tensions between privacy and ubiquity, resource management and performance optimization, and access and monopolization. Increased academic focus on these practitioners can facilitate a more holistic understanding of ML limitations, and so is useful for prescribing a research agenda to facilitate responsible ML development for all.

Post-hoc explanation methods are gaining popularity as tools for interpreting, understanding, and debugging neural networks. Most post-hoc methods explain decisions in response to individual inputs. These individual inputs are typically drawn from the test set; however, the test set may be biased or may only sparsely invoke some model behaviours. To address these challenges, we introduce Bayes-TrEx, a model-agnostic method for generating distribution-conforming examples of known prediction confidence. Using a classifier prediction and a data generator, Bayes-TrEx can be used to visualize class boundaries; to find in-distribution adversarial examples; to understand novel-class extrapolation; and to expose neural network overconfidence. We demonstrate Bayes-TrEx with rendered data (CLEVR) and organic data (MNIST, Fashion-MNIST). Code: this http URL.

Though neural network models demonstrate impressive performance, we do not understand exactly how these black-box models make individual predictions. This drawback has led to substantial research devoted to understand these models in areas such as robustness, interpretability, and generalization ability. In this paper, we consider the problem of exploring the prediction level sets of a classifier using probabilistic programming. We define a prediction level set to be the set of examples for which the predictor has the same specified prediction confidence with respect to some arbitrary data distribution. Notably, our sampling-based method does not require the classifier to be differentiable, making it compatible with arbitrary classifiers. As a specific instantiation, if we take the classifier to be a neural network and the data distribution to be that of the training data, we can obtain examples that will result in specified predictions by the neural network. We demonstrate this technique with experiments on a synthetic dataset and MNIST. Such level sets in classification may facilitate human understanding of classification behaviors.

Knowledge compilation techniques translate propositional theories into equivalent forms to increase their computational tractability. But, how should we best present these propositional theories to a human? We analyze the standard taxonomy of propositional theories for relative interpretability across three model domains: highway driving, emergency triage, and the chopsticks game. We generate decision-making agents which produce logical explanations for their actions and apply knowledge compilation to these explanations. Then, we evaluate how quickly, accurately, and confidently users comprehend the generated explanations. We find that domain, formula size, and negated logical connectives significantly affect comprehension while formula properties typically associated with interpretability are not strong predictors of human ability to comprehend the theory.