To navigate safely and efficiently in crowded spaces, robots should not only perceive the current state of the environment but also anticipate future human movements. In this paper, we propose a reinforcement learning architecture, namely Falcon, to tackle socially-aware navigation by explicitly predicting human trajectories and penalizing actions that block future human paths. To facilitate realistic evaluation, we introduce a novel SocialNav benchmark containing two new datasets, Social-HM3D and Social-MP3D. This benchmark offers large-scale photo-realistic indoor scenes populated with a reasonable amount of human agents based on scene area size, incorporating natural human movements and trajectory patterns. We conduct a detailed experimental analysis with the state-of-the-art learning-based method and two classic rule-based path-planning algorithms on the new benchmark. The results demonstrate the importance of future prediction and our method achieves the best task success rate of 55% while maintaining about 90% personal space compliance. We will release our code and datasets. Videos of demonstrations can be viewed at https://zeying-gong.github.io/projects/falcon/ .

In image-based reinforcement learning (RL), policies usually operate in two steps: first extracting lower-dimensional features from raw images (the "recognition" step), and then taking actions based on the extracted features (the "decision" step). Extracting features that are spuriously correlated with performance or irrelevant for decision-making can lead to poor generalization performance, known as observational overfitting in image-based RL. In such cases, it can be hard to quantify how much of the error can be attributed to poor feature extraction vs. poor decision-making. In order to disentangle the two sources of error, we introduce the notions of recognition regret and decision regret. Using these notions, we characterize and disambiguate the two distinct causes behind observational overfitting: over-specific representations, which include features that are not needed for optimal decision-making (leading to high decision regret), vs. under-specific representations, which only include a limited set of features that were spuriously correlated with performance during training (leading to high recognition regret). Finally, we provide illustrative examples of observational overfitting due to both over-specific and under-specific representations in maze environments as well as the Atari game Pong.

Building on the foundations of our previous work, this paper introduces Arena 4.0, a significant advancement over Arena 3.0, Arena-Bench, Arena 1.0, and Arena 2.0. Arena 4.0 offers three key novel contributions: (1) a generative-model-based world and scenario generation approach that utilizes large language models (LLMs) and diffusion models to dynamically generate complex, human-centric environments from text prompts or 2D floorplans, useful for the development and benchmarking of social navigation strategies; (2) a comprehensive 3D model database, extendable with additional 3D assets that are semantically linked and annotated for dynamic spawning and arrangement within 3D worlds; and (3) a complete migration to ROS 2, enabling compatibility with modern hardware and enhanced functionalities for improved navigation, usability, and easier deployment on real robots. We evaluated the platform's performance through a comprehensive user study, demonstrating significant improvements in usability and efficiency compared to previous versions. Arena 4.0 is openly available at https://github.com/Arena-Rosnav.

This paper introduces a novel framework combining LLM agents as proxies for human strategic behavior with reinforcement learning (RL) to engage these agents in evolving strategic interactions within team environments. Our approach extends traditional agent-based simulations by using strategic LLM agents (SLA) and introducing dynamic and adaptive governance through a pro-social promoting RL agent (PPA) that modulates information access across agents in a network, optimizing social welfare and promoting pro-social behavior. Through validation in iterative games, including the prisoner's dilemma, we demonstrate that SLA agents exhibit nuanced strategic adaptations. The PPA agent effectively learns to adjust information transparency, resulting in enhanced cooperation rates. This framework offers significant insights into AI-mediated social dynamics, contributing to the deployment of AI in real-world team settings.

This paper illustrates the central role of loss functions in data-driven decision making, providing a comprehensive survey on their influence in cost-sensitive classification (CSC) and reinforcement learning (RL). We demonstrate how different regression loss functions affect the sample efficiency and adaptivity of value-based decision making algorithms. Across multiple settings, we prove that algorithms using the binary cross-entropy loss achieve first-order bounds scaling with the optimal policy's cost and are much more efficient than the commonly used squared loss. Moreover, we prove that distributional algorithms using the maximum likelihood loss achieve second-order bounds scaling with the policy variance and are even sharper than first-order bounds. This in particular proves the benefits of distributional RL. We hope that this paper serves as a guide analyzing decision making algorithms with varying loss functions, and can inspire the reader to seek out better loss functions to improve any decision making algorithm.

We want to enable fine manipulation with a multi-fingered robotic hand by using modern deep reinforcement learning methods. Key for fine manipulation is a spatially resolved tactile sensor. Here, we present a novel model of a tactile skin that can be used together with rigid-body (hence fast) physics simulators. The model considers the softness of the real fingertips such that a contact can spread across multiple taxels of the sensor depending on the contact geometry. We calibrate the model parameters to allow for an accurate simulation of the real-world sensor. For this, we present a self-contained calibration method without external tools or sensors. To demonstrate the validity of our approach, we learn two challenging fine manipulation tasks: Rolling a marble and a bolt between two fingers. We show in simulation experiments that tactile feedback is crucial for precise manipulation and reaching sub-taxel resolution of < 1 mm (despite a taxel spacing of 4 mm). Moreover, we demonstrate that all policies successfully transfer from the simulation to the real robotic hand.

Recently, large language model (LLM)-based agents have made significant advances across various fields. One of the most popular research areas involves applying these agents to video games. Traditionally, these methods have relied on game APIs to access in-game environmental and action data. However, this approach is limited by the availability of APIs and does not reflect how humans play games. With the advent of vision language models (VLMs), agents now have enhanced visual understanding capabilities, enabling them to interact with games using only visual inputs. Despite these advances, current approaches still face challenges in action-oriented tasks, particularly in action role-playing games (ARPGs), where reinforcement learning methods are prevalent but suffer from poor generalization and require extensive training. To address these limitations, we select an ARPG, ``Black Myth: Wukong'', as a research platform to explore the capability boundaries of existing VLMs in scenarios requiring visual-only input and complex action output. We define 12 tasks within the game, with 75% focusing on combat, and incorporate several state-of-the-art VLMs into this benchmark. Additionally, we will release a human operation dataset containing recorded gameplay videos and operation logs, including mouse and keyboard actions. Moreover, we propose a novel VARP (Vision Action Role-Playing) agent framework, consisting of an action planning system and a visual trajectory system. Our framework demonstrates the ability to perform basic tasks and succeed in 90% of easy and medium-level combat scenarios. This research aims to provide new insights and directions for applying multimodal agents in complex action game environments. The code and datasets will be made available at https://varp-agent.github.io/.

Robust policies enable reinforcement learning agents to effectively adapt to and operate in unpredictable, dynamic, and ever-changing real-world environments. Factored representations, which break down complex state and action spaces into distinct components, can improve generalization and sample efficiency in policy learning. In this paper, we explore how the curriculum of an agent using a factored state representation affects the robustness of the learned policy. We experimentally demonstrate three simple curricula, such as varying only the variable of highest regret between episodes, that can significantly enhance policy robustness, offering practical insights for reinforcement learning in complex environments.

In multi-robot systems, relative localization between platforms plays a crucial role in many tasks, such as leader following, target tracking, or cooperative maneuvering. State of the Art (SotA) approaches either rely on infrastructure-based or on infrastructure-less setups. The former typically achieve high localization accuracy but require fixed external structures. The latter provide more flexibility, however, most of the works use cameras or lidars that require Line-of-Sight (LoS) to operate. Ultra Wide Band (UWB) devices are emerging as a viable alternative to build infrastructure-less solutions that do not require LoS. These approaches directly deploy the UWB sensors on the robots. However, they require that at least one of the platforms is static, limiting the advantages of an infrastructure-less setup. In this work, we remove this constraint and introduce an active method for infrastructure-less relative localization. Our approach allows the robot to adapt its position to minimize the relative localization error of the other platform. To this aim, we first design a specialized anchor placement for the active localization task. Then, we propose a novel UWB Relative Localization Loss that adapts the Geometric Dilution Of Precision metric to the infrastructure-less scenario. Lastly, we leverage this loss function to train an active Deep Reinforcement Learning-based controller for UWB relative localization. An extensive simulation campaign and real-world experiments validate our method, showing up to a 60% reduction of the localization error compared to current SotA approaches.

Stress testing is an approach for evaluating the reliability of systems under extreme conditions which help reveal vulnerable scenarios that standard testing may overlook. Identifying such scenarios is of great importance in autonomous vehicles (AV) and other safety-critical systems. Since failure events are rare, naive random search approaches require a large number of vehicle operation hours to identify potential system failures. Adaptive Stress Testing (AST) is a method addressing this constraint by effectively exploring the failure trajectories of AV using a Markov decision process and employs reinforcement learning techniques to identify driving scenarios with high probability of failures. However, existing AST frameworks are able to handle only simple scenarios, such as one vehicle moving longitudinally on a single lane road which is not realistic and has a limited applicability. In this paper, we propose a novel AST framework to systematically explore corner cases of intelligent driving models that can result in safety concerns involving both longitudinal and lateral vehicle's movements. Specially, we develop a new reward function for Deep Reinforcement Learning to guide the AST in identifying crash scenarios based on the collision probability estimate between the AV under test (i.e., the ego vehicle) and the trajectory of other vehicles on the multi-lane roads. To demonstrate the effectiveness of our framework, we tested it with a complex driving model vehicle that can be controlled in both longitudinal and lateral directions. Quantitative and qualitative analyses of our experimental results demonstrate that our framework outperforms the state-of-the-art AST scheme in identifying corner cases with complex driving maneuvers.