Erdem Bıyık

Stephen Casper, Xander Davies, Claudia Shi et al. · berkeley, eth-zurich

Reinforcement learning from human feedback (RLHF) is a technique for training AI systems to align with human goals. RLHF has emerged as the central method used to finetune state-of-the-art large language models (LLMs). Despite this popularity, there has been relatively little public work systematizing its flaws. In this paper, we (1) survey open problems and fundamental limitations of RLHF and related methods; (2) overview techniques to understand, improve, and complement RLHF in practice; and (3) propose auditing and disclosure standards to improve societal oversight of RLHF systems. Our work emphasizes the limitations of RLHF and highlights the importance of a multi-faceted approach to the development of safer AI systems.

Sumedh A Sontakke, Jesse Zhang, Sébastien M. R. Arnold et al.

Reward specification is a notoriously difficult problem in reinforcement learning, requiring extensive expert supervision to design robust reward functions. Imitation learning (IL) methods attempt to circumvent these problems by utilizing expert demonstrations but typically require a large number of in-domain expert demonstrations. Inspired by advances in the field of Video-and-Language Models (VLMs), we present RoboCLIP, an online imitation learning method that uses a single demonstration (overcoming the large data requirement) in the form of a video demonstration or a textual description of the task to generate rewards without manual reward function design. Additionally, RoboCLIP can also utilize out-of-domain demonstrations, like videos of humans solving the task for reward generation, circumventing the need to have the same demonstration and deployment domains. RoboCLIP utilizes pretrained VLMs without any finetuning for reward generation. Reinforcement learning agents trained with RoboCLIP rewards demonstrate 2-3 times higher zero-shot performance than competing imitation learning methods on downstream robot manipulation tasks, doing so using only one video/text demonstration.

Vivek Myers, Erdem Bıyık, Dorsa Sadigh

Robot policies need to adapt to human preferences and/or new environments. Human experts may have the domain knowledge required to help robots achieve this adaptation. However, existing works often require costly offline re-training on human feedback, and those feedback usually need to be frequent and too complex for the humans to reliably provide. To avoid placing undue burden on human experts and allow quick adaptation in critical real-world situations, we propose designing and sparingly presenting easy-to-answer pairwise action preference queries in an online fashion. Our approach designs queries and determines when to present them to maximize the expected value derived from the queries' information. We demonstrate our approach with experiments in simulation, human user studies, and real robot experiments. In these settings, our approach outperforms baseline techniques while presenting fewer queries to human experts. Experiment videos, code and appendices are found at https://sites.google.com/view/onlineactivepreferences.

Zhangjie Cao, Erdem Bıyık, Guy Rosman et al.

Multi-agent interactions are important to model for forecasting other agents' behaviors and trajectories. At a certain time, to forecast a reasonable future trajectory, each agent needs to pay attention to the interactions with only a small group of most relevant agents instead of unnecessarily paying attention to all the other agents. However, existing attention modeling works ignore that human attention in driving does not change rapidly, and may introduce fluctuating attention across time steps. In this paper, we formulate an attention model for multi-agent interactions based on a total variation temporal smoothness prior and propose a trajectory prediction architecture that leverages the knowledge of these attended interactions. We demonstrate how the total variation attention prior along with the new sequence prediction loss terms leads to smoother attention and more sample-efficient learning of multi-agent trajectory prediction, and show its advantages in terms of prediction accuracy by comparing it with the state-of-the-art approaches on both synthetic and naturalistic driving data. We demonstrate the performance of our algorithm for trajectory prediction on the INTERACTION dataset on our website.

Erdem Bıyık

When robots enter everyday human environments, they need to understand their tasks and how they should perform those tasks. To encode these, reward functions, which specify the objective of a robot, are employed. However, designing reward functions can be extremely challenging for complex tasks and environments. A promising approach is to learn reward functions from humans. Recently, several robot learning works embrace this approach and leverage human demonstrations to learn the reward functions. Known as inverse reinforcement learning, this approach relies on a fundamental assumption: humans can provide near-optimal demonstrations to the robot. Unfortunately, this is rarely the case: human demonstrations to the robot are often suboptimal due to various reasons, e.g., difficulty of teleoperation, robot having high degrees of freedom, or humans' cognitive limitations. This thesis is an attempt towards learning reward functions from human users by using other, more reliable data modalities. Specifically, we study how reward functions can be learned using comparative feedback, in which the human user compares multiple robot trajectories instead of (or in addition to) providing demonstrations. To this end, we first propose various forms of comparative feedback, e.g., pairwise comparisons, best-of-many choices, rankings, scaled comparisons; and describe how a robot can use these various forms of human feedback to infer a reward function, which may be parametric or non-parametric. Next, we propose active learning techniques to enable the robot to ask for comparison feedback that optimizes for the expected information that will be gained from that user feedback. Finally, we demonstrate the applicability of our methods in a wide variety of domains, ranging from autonomous driving simulations to home robotics, from standard reinforcement learning benchmarks to lower-body exoskeletons.

Rongtao Zhang, Xin Zhu, Masoume Pourebadi Khotbehsara et al.

Individual differences in vibrotactile perception underscore the growing importance of personalization as haptic feedback becomes more prevalent in interactive systems. We propose Vibrotactile Preference Learning (VPL), a system that captures user-specific preference spaces over vibrotactile parameters via Gaussian-process-based uncertainty-aware preference learning. VPL uses an expected information gain-based acquisition strategy to guide query selection over 40 rounds of pairwise comparisons of overall user preference, augmented with user-reported uncertainty, enabling efficient exploration of the parameter space. We evaluate VPL in a user study (N = 13) using the vibrotactile feedback from a Microsoft Xbox controller, showing that it efficiently learns individualized preferences while maintaining comfortable, low-workload user interactions. These results highlight the potential of VPL for scalable personalization of vibrotactile experiences.

Nathaniel Haynam, Adam Khoja, Dhruv Kumar et al.

When reward functions are hand-designed, deep reinforcement learning algorithms often suffer from reward misspecification, causing them to learn suboptimal policies in terms of the intended task objectives. In the single-agent case, inverse reinforcement learning (IRL) techniques attempt to address this issue by inferring the reward function from expert demonstrations. However, in multi-agent problems, misalignment between the learned and true objectives is exacerbated due to increased environment non-stationarity and variance that scales with multiple agents. As such, in multi-agent general-sum games, multi-agent IRL algorithms have difficulty balancing cooperative and competitive objectives. To address these issues, we propose Multi-Agent Marginal Q-Learning from Demonstrations (MAMQL), a novel sample-efficient framework for multi-agent IRL. For each agent, MAMQL learns a critic marginalized over the other agents' policies, allowing for a well-motivated use of Boltzmann policies in the multi-agent context. We identify a connection between optimal marginalized critics and single-agent soft-Q IRL, allowing us to apply a direct, simple optimization criterion from the single-agent domain. Across our experiments on three different simulated domains, MAMQL significantly outperforms previous multi-agent methods in average reward, sample efficiency, and reward recovery by often more than 2-5x. We make our code available at https://sites.google.com/view/mamql .

Minjune Hwang, Yigit Korkmaz, Daniel Seita et al.

Preference-based reward learning is widely used for shaping agent behavior to match a user's preference, yet its sparse binary feedback makes it especially vulnerable to causal confusion. The learned reward often latches onto spurious features that merely co-occur with preferred trajectories during training, collapsing when those correlations disappear or reverse at test time. We introduce ReCouPLe, a lightweight framework that uses natural language rationales to provide the missing causal signal. Each rationale is treated as a guiding projection axis in an embedding space, training the model to score trajectories based on features aligned with that axis while de-emphasizing context that is unrelated to the stated reason. Because the same rationales (e.g., "avoids collisions", "completes the task faster") can appear across multiple tasks, ReCouPLe naturally reuses the same causal direction whenever tasks share semantics, and transfers preference knowledge to novel tasks without extra data or language-model fine-tuning. Our learned reward model can ground preferences on the articulated reason, aligning better with user intent and generalizing beyond spurious features. ReCouPLe outperforms baselines by up to 1.5x in reward accuracy under distribution shifts, and 2x in downstream policy performance in novel tasks. We have released our code at https://github.com/mj-hwang/ReCouPLe

Yigit Korkmaz, Urvi Bhuwania, Ayush Jain et al.

Value-based algorithms are a cornerstone of off-policy reinforcement learning due to their simplicity and training stability. However, their use has traditionally been restricted to discrete action spaces, as they rely on estimating Q-values for individual state-action pairs. In continuous action spaces, evaluating the Q-value over the entire action space becomes computationally infeasible. To address this, actor-critic methods are typically employed, where a critic is trained on off-policy data to estimate Q-values, and an actor is trained to maximize the critic's output. Despite their popularity, these methods often suffer from instability during training. In this work, we propose a purely value-based framework for continuous control that revisits structural maximization of Q-functions, introducing a set of key architectural and algorithmic choices to enable efficient and stable learning. We evaluate the proposed actor-free Q-learning approach on a range of standard simulation tasks, demonstrating performance and sample efficiency on par with state-of-the-art baselines, without the cost of learning a separate actor. Particularly, in environments with constrained action spaces, where the value functions are typically non-smooth, our method with structural maximization outperforms traditional actor-critic methods with gradient-based maximization. We have released our code at https://github.com/USC-Lira/Q3C.

Zhaoyang Li, Zhan Ling, Yuchen Zhou et al.

Large Vision-Language Models (LVLMs) excel at captioning, visual question answering, and robotics by combining vision and language, yet they often miss obvious objects or hallucinate nonexistent ones in atypical scenes. We examine these failures through the lens of uncertainty, focusing on contextual incongruity, where objects appear unexpectedly or fail to appear in expected contexts, and show that such cases increase recognition difficulty for state-of-the-art LVLMs. To study this regime, we introduce the Object Recognition in Incongruous Context (ORIC) framework, which constructs incongruous object-context pairs through two complementary strategies: (1) LLM-guided sampling to identify hard-to-recognize objects present in the image and (2) CLIP-guided sampling to mine plausible but absent ones. Applied to MSCOCO, ORIC produces ORIC-Bench and ORIC-style training data. Evaluating 18 LVLMs and 2 open-vocabulary detectors reveals substantial performance drops and bias patterns under incongruous contexts. Fine-tuning Qwen3-VL-8B-Instruct with Visual Reinforcement Fine-Tuning on 600 ORIC-style samples improves results on ORIC-Bench, AMBER, and HallusionBench. Overall, we show that contextual incongruity is a key source of uncertainty and provide tools for more reliable LVLMs. The code is available at https://github.com/ZhaoyangLi-1/ORIC.

Erdem Bıyık, Aditi Talati, Dorsa Sadigh

Reward learning is a fundamental problem in human-robot interaction to have robots that operate in alignment with what their human user wants. Many preference-based learning algorithms and active querying techniques have been proposed as a solution to this problem. In this paper, we present APReL, a library for active preference-based reward learning algorithms, which enable researchers and practitioners to experiment with the existing techniques and easily develop their own algorithms for various modules of the problem. APReL is available at https://github.com/Stanford-ILIAD/APReL.

An-Chieh Cheng, Yandong Ji, Zhaojing Yang et al.

This paper proposes to solve the problem of Vision-and-Language Navigation with legged robots, which not only provides a flexible way for humans to command but also allows the robot to navigate through more challenging and cluttered scenes. However, it is non-trivial to translate human language instructions all the way to low-level leg joint actions. We propose NaVILA, a 2-level framework that unifies a Vision-Language-Action model (VLA) with locomotion skills. Instead of directly predicting low-level actions from VLA, NaVILA first generates mid-level actions with spatial information in the form of language, (e.g., "moving forward 75cm"), which serves as an input for a visual locomotion RL policy for execution. NaVILA substantially improves previous approaches on existing benchmarks. The same advantages are demonstrated in our newly developed benchmarks with IsaacLab, featuring more realistic scenes, low-level controls, and real-world robot experiments. We show more results at https://navila-bot.github.io/

Erdem Bıyık, Nima Anari, Dorsa Sadigh

Data generation and labeling are often expensive in robot learning. Preference-based learning is a concept that enables reliable labeling by querying users with preference questions. Active querying methods are commonly employed in preference-based learning to generate more informative data at the expense of parallelization and computation time. In this paper, we develop a set of novel algorithms, batch active preference-based learning methods, that enable efficient learning of reward functions using as few data samples as possible while still having short query generation times and also retaining parallelizability. We introduce a method based on determinantal point processes (DPP) for active batch generation and several heuristic-based alternatives. Finally, we present our experimental results for a variety of robotics tasks in simulation. Our results suggest that our batch active learning algorithm requires only a few queries that are computed in a short amount of time. We showcase one of our algorithms in a study to learn human users' preferences.

Anthony Liang, Jesse Thomason, Erdem Bıyık · uw

Training robots to perform complex control tasks from high-dimensional pixel input using reinforcement learning (RL) is sample-inefficient, because image observations are comprised primarily of task-irrelevant information. By contrast, humans are able to visually attend to task-relevant objects and areas. Based on this insight, we introduce Visual Saliency-Guided Reinforcement Learning (ViSaRL). Using ViSaRL to learn visual representations significantly improves the success rate, sample efficiency, and generalization of an RL agent on diverse tasks including DeepMind Control benchmark, robot manipulation in simulation and on a real robot. We present approaches for incorporating saliency into both CNN and Transformer-based encoders. We show that visual representations learned using ViSaRL are robust to various sources of visual perturbations including perceptual noise and scene variations. ViSaRL nearly doubles success rate on the real-robot tasks compared to the baseline which does not use saliency.

Yigit Korkmaz, Erdem Bıyık

Imitation learning techniques have been shown to be highly effective in real-world control scenarios, such as robotics. However, these approaches not only suffer from compounding error issues but also require human experts to provide complete trajectories. Although there exist interactive methods where an expert oversees the robot and intervenes if needed, these extensions usually only utilize the data collected during intervention periods and ignore the feedback signal hidden in non-intervention timesteps. In this work, we create a model to formulate how the interventions occur in such cases, and show that it is possible to learn a policy with just a handful of expert interventions. Our key insight is that it is possible to get crucial information about the quality of the current state and the optimality of the chosen action from expert feedback, regardless of the presence or the absence of intervention. We evaluate our method on various discrete and continuous simulation environments, a real-world robotic manipulation task, as well as a human subject study. Videos and the code can be found at https://liralab.usc.edu/mile .

Evan Ellis, Gaurav R. Ghosal, Stuart J. Russell et al.

Preference-based reward learning is a popular technique for teaching robots and autonomous systems how a human user wants them to perform a task. Previous works have shown that actively synthesizing preference queries to maximize information gain about the reward function parameters improves data efficiency. The information gain criterion focuses on precisely identifying all parameters of the reward function. This can potentially be wasteful as many parameters may result in the same reward, and many rewards may result in the same behavior in the downstream tasks. Instead, we show that it is possible to optimize for learning the reward function up to a behavioral equivalence class, such as inducing the same ranking over behaviors, distribution over choices, or other related definitions of what makes two rewards similar. We introduce a tractable framework that can capture such definitions of similarity. Our experiments in a synthetic environment, an assistive robotics environment with domain transfer, and a natural language processing problem with real datasets demonstrate the superior performance of our querying method over the state-of-the-art information gain method.

Yiyang Ling, Karan Owalekar, Oluwatobiloba Adesanya et al.

Motion planning involves determining a sequence of robot configurations to reach a desired pose, subject to movement and safety constraints. Traditional motion planning finds collision-free paths, but this is overly restrictive in clutter, where it may not be possible for a robot to accomplish a task without contact. In addition, contacts range from relatively benign (e.g. brushing a soft pillow) to more dangerous (e.g. toppling a glass vase), making it difficult to characterize which may be acceptable. In this paper, we propose IMPACT, a novel motion planning framework that uses Vision-Language Models (VLMs) to infer environment semantics, identifying which parts of the environment can best tolerate contact based on object properties and locations. Our approach generates an anisotropic cost map that encodes directional push safety. We pair this map with a contact-aware A* planner to find stable contact-rich paths. We perform experiments using 20 simulation and 10 real-world scenes and assess using task success rate, object displacements, and feedback from human evaluators. Our results over 3200 simulation and 200 real-world trials suggest that IMPACT enables efficient contact-rich motion planning in cluttered settings while outperforming alternative methods and ablations. Our project website is available at https://impact-planning.github.io/.

Anthony Liang, Guy Tennenholtz, Chih-wei Hsu et al.

We introduce DynaMITE-RL, a meta-reinforcement learning (meta-RL) approach to approximate inference in environments where the latent state evolves at varying rates. We model episode sessions - parts of the episode where the latent state is fixed - and propose three key modifications to existing meta-RL methods: consistency of latent information within sessions, session masking, and prior latent conditioning. We demonstrate the importance of these modifications in various domains, ranging from discrete Gridworld environments to continuous-control and simulated robot assistive tasks, demonstrating that DynaMITE-RL significantly outperforms state-of-the-art baselines in sample efficiency and inference returns.

Amin Banayeeanzade, Fatemeh Bahrani, Yutai Zhou et al.

Imitation Learning (IL) is a widely adopted approach which enables agents to learn from human expert demonstrations by framing the task as a supervised learning problem. However, IL often suffers from causal confusion, where agents misinterpret spurious correlations as causal relationships, leading to poor performance in testing environments with distribution shift. To address this issue, we introduce GAze-Based Regularization in Imitation Learning (GABRIL), a novel method that leverages the human gaze data gathered during the data collection phase to guide the representation learning in IL. GABRIL utilizes a regularization loss which encourages the model to focus on causally relevant features identified through expert gaze and consequently mitigates the effects of confounding variables. We validate our approach in Atari environments and the Bench2Drive benchmark in CARLA by collecting human gaze datasets and applying our method in both domains. Experimental results show that the improvement of GABRIL over behavior cloning is around 179% more than the same number for other baselines in the Atari and 76% in the CARLA setup. Finally, we show that our method provides extra explainability when compared to regular IL agents.

Harbani Jaggi, Kashyap Murali, Eve Fleisig et al.

When annotators disagree, predicting the labels given by individual annotators can capture nuances overlooked by traditional label aggregation. We introduce three approaches to predicting individual annotator ratings on the toxicity of text by incorporating individual annotator-specific information: a neural collaborative filtering (NCF) approach, an in-context learning (ICL) approach, and an intermediate embedding-based architecture. We also study the utility of demographic information for rating prediction. NCF showed limited utility; however, integrating annotator history, demographics, and survey information permits both the embedding-based architecture and ICL to substantially improve prediction accuracy, with the embedding-based architecture outperforming the other methods. We also find that, if demographics are predicted from survey information, using these imputed demographics as features performs comparably to using true demographic data. This suggests that demographics may not provide substantial information for modeling ratings beyond what is captured in survey responses. Our findings raise considerations about the relative utility of different types of annotator information and provide new approaches for modeling annotators in subjective NLP tasks.

Ayush Jain, Norio Kosaka, Xinhu Li et al.

In reinforcement learning, off-policy actor-critic methods like DDPG and TD3 use deterministic policy gradients: the Q-function is learned from environment data, while the actor maximizes it via gradient ascent. We observe that in complex tasks such as dexterous manipulation and restricted locomotion with mobility constraints, the Q-function exhibits many local optima, making gradient ascent prone to getting stuck. To address this, we introduce SAVO, an actor architecture that (i) generates multiple action proposals and selects the one with the highest Q-value, and (ii) approximates the Q-function repeatedly by truncating poor local optima to guide gradient ascent more effectively. We evaluate tasks such as restricted locomotion, dexterous manipulation, and large discrete-action space recommender systems and show that our actor finds optimal actions more frequently and outperforms alternate actor architectures.

Litian Gong, Fatemeh Bahrani, Yutai Zhou et al.

AutoFocus-IL is a simple yet effective method to improve data efficiency and generalization in visual imitation learning by guiding policies to attend to task-relevant features rather than distractors and spurious correlations. Although saliency regularization has emerged as a promising way to achieve this, existing approaches typically require costly supervision such as human gaze data or manual saliency annotations. In contrast, AutoFocus-IL leverages vision-language models (VLMs) to automatically identify and track key objects in demonstrations, generating temporal saliency maps that highlight causal visual signals while suppressing distractors. These maps are then used to regularize behavior cloning policies, yielding stronger alignment between visual attention and task-relevant cues. Experiments in both the CARLA simulator and real-robot manipulation tasks demonstrate that AutoFocus-IL not only outperforms standard behavior cloning but also surpasses state-of-the-art baselines that assume privileged access to human supervision, such as gaze data. Code, datasets, and trained policy videos are available at https://AutoFocus-IL.github.io/.

Xinhu Li, Ayush Jain, Zhaojing Yang et al.

Learning from demonstrations enables experts to teach robots complex tasks using interfaces such as kinesthetic teaching, joystick control, and sim-to-real transfer. However, these interfaces often constrain the expert's ability to demonstrate optimal behavior due to indirect control, setup restrictions, and hardware safety. For example, a joystick can move a robotic arm only in a 2D plane, even though the robot operates in a higher-dimensional space. As a result, the demonstrations collected by constrained experts lead to suboptimal performance of the learned policies. This raises a key question: Can a robot learn a better policy than the one demonstrated by a constrained expert? We address this by allowing the agent to go beyond direct imitation of expert actions and explore shorter and more efficient trajectories. We use the demonstrations to infer a state-only reward signal that measures task progress, and self-label reward for unknown states using temporal interpolation. Our approach outperforms common imitation learning in both sample efficiency and task completion time. On a real WidowX robotic arm, it completes the task in 12 seconds, 10x faster than behavioral cloning, as shown in real-robot videos on https://sites.google.com/view/constrainedexpert .

Jesse Zhang, Marius Memmel, Kevin Kim et al.

Robotic manipulation policies often fail to generalize because they must simultaneously learn where to attend, what actions to take, and how to execute them. We argue that high-level reasoning about where and what can be offloaded to vision-language models (VLMs), leaving policies to specialize in how to act. We present PEEK (Policy-agnostic Extraction of Essential Keypoints), which fine-tunes VLMs to predict a unified point-based intermediate representation: 1. end-effector paths specifying what actions to take, and 2. task-relevant masks indicating where to focus. These annotations are directly overlaid onto robot observations, making the representation policy-agnostic and transferable across architectures. To enable scalable training, we introduce an automatic annotation pipeline, generating labeled data across 20+ robot datasets spanning 9 embodiments. In real-world evaluations, PEEK consistently boosts zero-shot generalization, including a 41.4x real-world improvement for a 3D policy trained only in simulation, and 2-3.5x gains for both large VLAs and small manipulation policies. By letting VLMs absorb semantic and visual complexity, PEEK equips manipulation policies with the minimal cues they need--where, what, and how. Website at https://peek-robot.github.io/.

Yimin Tang, Xiao Xiong, Jingyi Xi et al.

Multi-Agent Path Finding (MAPF), which focuses on finding collision-free paths for multiple robots, is crucial for applications ranging from aerial swarms to warehouse automation. Solving MAPF is NP-hard so learning-based approaches for MAPF have gained attention, particularly those leveraging deep neural networks. Nonetheless, despite the community's continued efforts, all learning-based MAPF planners still rely on decentralized planning due to variability in the number of agents and map sizes. We have developed the first centralized learning-based policy for MAPF problem called RAILGUN. RAILGUN is not an agent-based policy but a map-based policy. By leveraging a CNN-based architecture, RAILGUN can generalize across different maps and handle any number of agents. We collect trajectories from rule-based methods to train our model in a supervised way. In experiments, RAILGUN outperforms most baseline methods and demonstrates great zero-shot generalization capabilities on various tasks, maps and agent numbers that were not seen in the training dataset.

Michelle Pan, Mariah Schrum, Vivek Myers et al.

Adaptive brain stimulation can treat neurological conditions such as Parkinson's disease and post-stroke motor deficits by influencing abnormal neural activity. Because of patient heterogeneity, each patient requires a unique stimulation policy to achieve optimal neural responses. Model-free reinforcement learning (MFRL) holds promise in learning effective policies for a variety of similar control tasks, but is limited in domains like brain stimulation by a need for numerous costly environment interactions. In this work we introduce Coprocessor Actor Critic, a novel, model-based reinforcement learning (MBRL) approach for learning neural coprocessor policies for brain stimulation. Our key insight is that coprocessor policy learning is a combination of learning how to act optimally in the world and learning how to induce optimal actions in the world through stimulation of an injured brain. We show that our approach overcomes the limitations of traditional MFRL methods in terms of sample efficiency and task success and outperforms baseline MBRL approaches in a neurologically realistic model of an injured brain.

Erdem Bıyık, Anusha Lalitha, Rajarshi Saha et al.

When humans collaborate with each other, they often make decisions by observing others and considering the consequences that their actions may have on the entire team, instead of greedily doing what is best for just themselves. We would like our AI agents to effectively collaborate in a similar way by capturing a model of their partners. In this work, we propose and analyze a decentralized Multi-Armed Bandit (MAB) problem with coupled rewards as an abstraction of more general multi-agent collaboration. We demonstrate that naïve extensions of single-agent optimal MAB algorithms fail when applied for decentralized bandit teams. Instead, we propose a Partner-Aware strategy for joint sequential decision-making that extends the well-known single-agent Upper Confidence Bound algorithm. We analytically show that our proposed strategy achieves logarithmic regret, and provide extensive experiments involving human-AI and human-robot collaboration to validate our theoretical findings. Our results show that the proposed partner-aware strategy outperforms other known methods, and our human subject studies suggest humans prefer to collaborate with AI agents implementing our partner-aware strategy.

Nils Wilde, Erdem Bıyık, Dorsa Sadigh et al.

Today's robots are increasingly interacting with people and need to efficiently learn inexperienced user's preferences. A common framework is to iteratively query the user about which of two presented robot trajectories they prefer. While this minimizes the users effort, a strict choice does not yield any information on how much one trajectory is preferred. We propose scale feedback, where the user utilizes a slider to give more nuanced information. We introduce a probabilistic model on how users would provide feedback and derive a learning framework for the robot. We demonstrate the performance benefit of slider feedback in simulations, and validate our approach in two user studies suggesting that scale feedback enables more effective learning in practice.

Vivek Myers, Erdem Bıyık, Nima Anari et al.

Learning from human feedback has shown to be a useful approach in acquiring robot reward functions. However, expert feedback is often assumed to be drawn from an underlying unimodal reward function. This assumption does not always hold including in settings where multiple experts provide data or when a single expert provides data for different tasks -- we thus go beyond learning a unimodal reward and focus on learning a multimodal reward function. We formulate the multimodal reward learning as a mixture learning problem and develop a novel ranking-based learning approach, where the experts are only required to rank a given set of trajectories. Furthermore, as access to interaction data is often expensive in robotics, we develop an active querying approach to accelerate the learning process. We conduct experiments and user studies using a multi-task variant of OpenAI's LunarLander and a real Fetch robot, where we collect data from multiple users with different preferences. The results suggest that our approach can efficiently learn multimodal reward functions, and improve data-efficiency over benchmark methods that we adapt to our learning problem.

Woodrow Z. Wang, Mark Beliaev, Erdem Bıyık et al.

Coordination is often critical to forming prosocial behaviors -- behaviors that increase the overall sum of rewards received by all agents in a multi-agent game. However, state of the art reinforcement learning algorithms often suffer from converging to socially less desirable equilibria when multiple equilibria exist. Previous works address this challenge with explicit reward shaping, which requires the strong assumption that agents can be forced to be prosocial. We propose using a less restrictive peer-rewarding mechanism, gifting, that guides the agents toward more socially desirable equilibria while allowing agents to remain selfish and decentralized. Gifting allows each agent to give some of their reward to other agents. We employ a theoretical framework that captures the benefit of gifting in converging to the prosocial equilibrium by characterizing the equilibria's basins of attraction in a dynamical system. With gifting, we demonstrate increased convergence of high risk, general-sum coordination games to the prosocial equilibrium both via numerical analysis and experiments.

Erdem Bıyık, Daniel A. Lazar, Ramtin Pedarsani et al.

Traffic congestion has large economic and social costs. The introduction of autonomous vehicles can potentially reduce this congestion by increasing road capacity via vehicle platooning and by creating an avenue for influencing people's choice of routes. We consider a network of parallel roads with two modes of transportation: (i) human drivers, who will choose the quickest route available to them, and (ii) a ride hailing service, which provides an array of autonomous vehicle route options, each with different prices, to users. We formalize a model of vehicle flow in mixed autonomy and a model of how autonomous service users make choices between routes with different prices and latencies. Developing an algorithm to learn the preferences of the users, we formulate a planning optimization that chooses prices to maximize a social objective. We demonstrate the benefit of the proposed scheme by comparing the results to theoretical benchmarks which we show can be efficiently calculated.

Sydney M. Katz, Amir Maleki, Erdem Bıyık et al.

Preference-based learning of reward functions, where the reward function is learned using comparison data, has been well studied for complex robotic tasks such as autonomous driving. Existing algorithms have focused on learning reward functions that are linear in a set of trajectory features. The features are typically hand-coded, and preference-based learning is used to determine a particular user's relative weighting for each feature. Designing a representative set of features to encode reward is challenging and can result in inaccurate models that fail to model the users' preferences or perform the task properly. In this paper, we present a method to learn both the relative weighting among features as well as additional features that help encode a user's reward function. The additional features are modeled as a neural network that is trained on the data from pairwise comparison queries. We apply our methods to a driving scenario used in previous work and compare the predictive power of our method to that of only hand-coded features. We perform additional analysis to interpret the learned features and examine the optimal trajectories. Our results show that adding an additional learned feature to the reward model enhances both its predictive power and expressiveness, producing unique results for each user.

Mark Beliaev, Erdem Bıyık, Daniel A. Lazar et al.

The COVID-19 pandemic has severely affected many aspects of people's daily lives. While many countries are in a re-opening stage, some effects of the pandemic on people's behaviors are expected to last much longer, including how they choose between different transport options. Experts predict considerably delayed recovery of the public transport options, as people try to avoid crowded places. In turn, significant increases in traffic congestion are expected, since people are likely to prefer using their own vehicles or taxis as opposed to riskier and more crowded options such as the railway. In this paper, we propose to use financial incentives to set the tradeoff between risk of infection and congestion to achieve safe and efficient transportation networks. To this end, we formulate a network optimization problem to optimize taxi fares. For our framework to be useful in various cities and times of the day without much designer effort, we also propose a data-driven approach to learn human preferences about transport options, which is then used in our taxi fare optimization. Our user studies and simulation experiments show our framework is able to minimize congestion and risk of infection.

Kejun Li, Maegan Tucker, Erdem Bıyık et al.

Characterizing what types of exoskeleton gaits are comfortable for users, and understanding the science of walking more generally, require recovering a user's utility landscape. Learning these landscapes is challenging, as walking trajectories are defined by numerous gait parameters, data collection from human trials is expensive, and user safety and comfort must be ensured. This work proposes the Region of Interest Active Learning (ROIAL) framework, which actively learns each user's underlying utility function over a region of interest that ensures safety and comfort. ROIAL learns from ordinal and preference feedback, which are more reliable feedback mechanisms than absolute numerical scores. The algorithm's performance is evaluated both in simulation and experimentally for three non-disabled subjects walking inside of a lower-body exoskeleton. ROIAL learns Bayesian posteriors that predict each exoskeleton user's utility landscape across four exoskeleton gait parameters. The algorithm discovers both commonalities and discrepancies across users' gait preferences and identifies the gait parameters that most influenced user feedback. These results demonstrate the feasibility of recovering gait utility landscapes from limited human trials.

Zheqing Zhu, Erdem Bıyık, Dorsa Sadigh

Multi-agent safe systems have become an increasingly important area of study as we can now easily have multiple AI-powered systems operating together. In such settings, we need to ensure the safety of not only each individual agent, but also the overall system. In this paper, we introduce a novel multi-agent safe learning algorithm that enables decentralized safe navigation when there are multiple different agents in the environment. This algorithm makes mild assumptions about other agents and is trained in a decentralized fashion, i.e. with very little prior knowledge about other agents' policies. Experiments show our algorithm performs well with the robots running other algorithms when optimizing various objectives.

Zhangjie Cao, Erdem Bıyık, Woodrow Z. Wang et al.

Autonomous driving has achieved significant progress in recent years, but autonomous cars are still unable to tackle high-risk situations where a potential accident is likely. In such near-accident scenarios, even a minor change in the vehicle's actions may result in drastically different consequences. To avoid unsafe actions in near-accident scenarios, we need to fully explore the environment. However, reinforcement learning (RL) and imitation learning (IL), two widely-used policy learning methods, cannot model rapid phase transitions and are not scalable to fully cover all the states. To address driving in near-accident scenarios, we propose a hierarchical reinforcement and imitation learning (H-ReIL) approach that consists of low-level policies learned by IL for discrete driving modes, and a high-level policy learned by RL that switches between different driving modes. Our approach exploits the advantages of both IL and RL by integrating them into a unified learning framework. Experimental results and user studies suggest our approach can achieve higher efficiency and safety compared to other methods. Analyses of the policies demonstrate our high-level policy appropriately switches between different low-level policies in near-accident driving situations.

Erdem Bıyık, Dylan P. Losey, Malayandi Palan et al.

Reward functions are a common way to specify the objective of a robot. As designing reward functions can be extremely challenging, a more promising approach is to directly learn reward functions from human teachers. Importantly, data from human teachers can be collected either passively or actively in a variety of forms: passive data sources include demonstrations, (e.g., kinesthetic guidance), whereas preferences (e.g., comparative rankings) are actively elicited. Prior research has independently applied reward learning to these different data sources. However, there exist many domains where multiple sources are complementary and expressive. Motivated by this general problem, we present a framework to integrate multiple sources of information, which are either passively or actively collected from human users. In particular, we present an algorithm that first utilizes user demonstrations to initialize a belief about the reward function, and then actively probes the user with preference queries to zero-in on their true reward. This algorithm not only enables us combine multiple data sources, but it also informs the robot when it should leverage each type of information. Further, our approach accounts for the human's ability to provide data: yielding user-friendly preference queries which are also theoretically optimal. Our extensive simulated experiments and user studies on a Fetch mobile manipulator demonstrate the superiority and the usability of our integrated framework.

Erdem Bıyık, Nicolas Huynh, Mykel J. Kochenderfer et al.

Designing reward functions is a challenging problem in AI and robotics. Humans usually have a difficult time directly specifying all the desirable behaviors that a robot needs to optimize. One common approach is to learn reward functions from collected expert demonstrations. However, learning reward functions from demonstrations introduces many challenges: some methods require highly structured models, e.g. reward functions that are linear in some predefined set of features, while others adopt less structured reward functions that on the other hand require tremendous amount of data. In addition, humans tend to have a difficult time providing demonstrations on robots with high degrees of freedom, or even quantifying reward values for given demonstrations. To address these challenges, we present a preference-based learning approach, where as an alternative, the human feedback is only in the form of comparisons between trajectories. Furthermore, we do not assume highly constrained structures on the reward function. Instead, we model the reward function using a Gaussian Process (GP) and propose a mathematical formulation to actively find a GP using only human preferences. Our approach enables us to tackle both inflexibility and data-inefficiency problems within a preference-based learning framework. Our results in simulations and a user study suggest that our approach can efficiently learn expressive reward functions for robotics tasks.

Erdem Bıyık, Malayandi Palan, Nicholas C. Landolfi et al.

Robots can learn the right reward function by querying a human expert. Existing approaches attempt to choose questions where the robot is most uncertain about the human's response; however, they do not consider how easy it will be for the human to answer! In this paper we explore an information gain formulation for optimally selecting questions that naturally account for the human's ability to answer. Our approach identifies questions that optimize the trade-off between robot and human uncertainty, and determines when these questions become redundant or costly. Simulations and a user study show our method not only produces easy questions, but also ultimately results in faster reward learning.

Daniel A. Lazar, Erdem Bıyık, Dorsa Sadigh et al.

Road congestion induces significant costs across the world, and road network disturbances, such as traffic accidents, can cause highly congested traffic patterns. If a planner had control over the routing of all vehicles in the network, they could easily reverse this effect. In a more realistic scenario, we consider a planner that controls autonomous cars, which are a fraction of all present cars. We study a dynamic routing game, in which the route choices of autonomous cars can be controlled and the human drivers react selfishly and dynamically. As the problem is prohibitively large, we use deep reinforcement learning to learn a policy for controlling the autonomous vehicles. This policy indirectly influences human drivers to route themselves in such a way that minimizes congestion on the network. To gauge the effectiveness of our learned policies, we establish theoretical results characterizing equilibria and empirically compare the learned policy results with best possible equilibria. We prove properties of equilibria on parallel roads and provide a polynomial-time optimization for computing the most efficient equilibrium. Moreover, we show that in the absence of these policies, high demand and network perturbations would result in large congestion, whereas using the policy greatly decreases the travel times by minimizing the congestion. To the best of our knowledge, this is the first work that employs deep reinforcement learning to reduce congestion by indirectly influencing humans' routing decisions in mixed-autonomy traffic.

Erdem Bıyık, Kenneth Wang, Nima Anari et al.

Data collection and labeling is one of the main challenges in employing machine learning algorithms in a variety of real-world applications with limited data. While active learning methods attempt to tackle this issue by labeling only the data samples that give high information, they generally suffer from large computational costs and are impractical in settings where data can be collected in parallel. Batch active learning methods attempt to overcome this computational burden by querying batches of samples at a time. To avoid redundancy between samples, previous works rely on some ad hoc combination of sample quality and diversity. In this paper, we present a new principled batch active learning method using Determinantal Point Processes, a repulsive point process that enables generating diverse batches of samples. We develop tractable algorithms to approximate the mode of a DPP distribution, and provide theoretical guarantees on the degree of approximation. We further demonstrate that an iterative greedy method for DPP maximization, which has lower computational costs but worse theoretical guarantees, still gives competitive results for batch active learning. Our experiments show the value of our methods on several datasets against state-of-the-art baselines.

Erdem Bıyık, Daniel A. Lazar, Dorsa Sadigh et al.

Autonomous vehicles have the potential to increase the capacity of roads via platooning, even when human drivers and autonomous vehicles share roads. However, when users of a road network choose their routes selfishly, the resulting traffic configuration may be very inefficient. Because of this, we consider how to influence human decisions so as to decrease congestion on these roads. We consider a network of parallel roads with two modes of transportation: (i) human drivers who will choose the quickest route available to them, and (ii) ride hailing service which provides an array of autonomous vehicle ride options, each with different prices, to users. In this work, we seek to design these prices so that when autonomous service users choose from these options and human drivers selfishly choose their resulting routes, road usage is maximized and transit delay is minimized. To do so, we formalize a model of how autonomous service users make choices between routes with different price/delay values. Developing a preference-based algorithm to learn the preferences of the users, and using a vehicle flow model related to the Fundamental Diagram of Traffic, we formulate a planning optimization to maximize a social objective and demonstrate the benefit of the proposed routing and learning scheme.

Erdem Bıyık, Jonathan Margoliash, Shahrouz Ryan Alimo et al.

We propose a safe exploration algorithm for deterministic Markov Decision Processes with unknown transition models. Our algorithm guarantees safety by leveraging Lipschitz-continuity to ensure that no unsafe states are visited during exploration. Unlike many other existing techniques, the provided safety guarantee is deterministic. Our algorithm is optimized to reduce the number of actions needed for exploring the safe space. We demonstrate the performance of our algorithm in comparison with baseline methods in simulation on navigation tasks.

Erdem Bıyık, Daniel Lazar, Ramtin Pedarsani et al.

Traffic congestion has large economic and social costs. The introduction of autonomous vehicles can potentially reduce this congestion, both by increasing network throughput and by enabling a social planner to incentivize users of autonomous vehicles to take longer routes that can alleviate congestion on more direct roads. We formalize the effects of altruistic autonomy on roads shared between human drivers and autonomous vehicles. In this work, we develop a formal model of road congestion on shared roads based on the fundamental diagram of traffic. We consider a network of parallel roads and provide algorithms that compute optimal equilibria that are robust to additional unforeseen demand. We further plan for optimal routings when users have varying degrees of altruism. We find that even with arbitrarily small altruism, total latency can be unboundedly better than without altruism, and that the best selfish equilibrium can be unboundedly better than the worst selfish equilibrium. We validate our theoretical results through microscopic traffic simulations and show average latency decrease of a factor of 4 from worst-case selfish equilibrium to the optimal equilibrium when autonomous vehicles are altruistic.

Erdem Bıyık, Dorsa Sadigh

Data generation and labeling are usually an expensive part of learning for robotics. While active learning methods are commonly used to tackle the former problem, preference-based learning is a concept that attempts to solve the latter by querying users with preference questions. In this paper, we will develop a new algorithm, batch active preference-based learning, that enables efficient learning of reward functions using as few data samples as possible while still having short query generation times. We introduce several approximations to the batch active learning problem, and provide theoretical guarantees for the convergence of our algorithms. Finally, we present our experimental results for a variety of robotics tasks in simulation. Our results suggest that our batch active learning algorithm requires only a few queries that are computed in a short amount of time. We then showcase our algorithm in a study to learn human users' preferences.