As language models (LMs) are widely utilized in personalized communication scenarios (e.g., sending emails, writing social media posts) and endowed with a certain level of agency, ensuring they act in accordance with the contextual privacy norms becomes increasingly critical. However, quantifying the privacy norm awareness of LMs and the emerging privacy risk in LM-mediated communication is challenging due to (1) the contextual and long-tailed nature of privacy-sensitive cases, and (2) the lack of evaluation approaches that capture realistic application scenarios. To address these challenges, we propose PrivacyLens, a novel framework designed to extend privacy-sensitive seeds into expressive vignettes and further into agent trajectories, enabling multi-level evaluation of privacy leakage in LM agents' actions. We instantiate PrivacyLens with a collection of privacy norms grounded in privacy literature and crowdsourced seeds. Using this dataset, we reveal a discrepancy between LM performance in answering probing questions and their actual behavior when executing user instructions in an agent setup. State-of-the-art LMs, like GPT-4 and Llama-3-70B, leak sensitive information in 25.68% and 38.69% of cases, even when prompted with privacy-enhancing instructions. We also demonstrate the dynamic nature of PrivacyLens by extending each seed into multiple trajectories to red-team LM privacy leakage risk. Dataset and code are available at https://github.com/SALT-NLP/PrivacyLens.

Fueled by advances in artificial intelligence, robotic process automation is impacting virtually every sector of the economy (McKinsey Global Institute 2017). The logistics sector lies at the core of this transformation: autonomous mobile robots are being deployed in tens of thousands of manufacturing and distribution facilities with a near-term $10-50 billion market potential (Grand View Research 2021, ABI Research 2021). A predominant operating model, shown in Figure 1, involves part-to-picker warehousing operations, which relies on robotic agents transporting shelves of inventory from a storage location to a workstation for a human operator to fulfill orders and back to a storage location. Robotic operations can improve throughput and working conditions by letting human workers focus on the more productive tasks, while improving system reliability. Yet, to truly take advantage of automation opportunities, modern warehousing systems require dedicated decision support tools to manage large robotic fleets and human-robot interactions in high-density operations. At the core of robotic process automation lies the computer vision, sensing, mapping and robotic technologies to empower autonomous agents--in our case, robots capable to move shelves of inventory. A subsequent problem involves control mechanisms to coordinate multiagent systems--in our case, to avoid conflicts and collisions between robots.

Consensus planning is a method for coordinating decision making across complex systems and organizations, including complex supply chain optimization pipelines. It arises when large interdependent distributed agents (systems) share common resources and must act in order to achieve a joint goal. In this paper, we introduce a generic Consensus Planning Protocol (CPP) to solve such problems. Our protocol allows for different agents to interact with the coordinating algorithm in different ways (e.g., as a primal or dual or proximal agent). In prior consensus planning work, all agents have been assumed to have the same interaction pattern (e.g., all dual agents or all primal agents or all proximal agents), most commonly using the Alternating Direction Method of Multipliers (ADMM) as proximal agents. However, this is often not a valid assumption in practice, where agents consist of large complex systems, and where we might not have the luxury of modifying these large complex systems at will. Our generic CPP allows for any mix of agents by combining ADMM-like updates for the proximal agents, dual ascent updates for the dual agents, and linearized ADMM updates for the primal agents. We prove convergence results for the generic CPP, namely a sublinear O(1/k) convergence rate under mild assumptions, and two-step linear convergence under stronger assumptions. We also discuss enhancements to the basic method and provide illustrative empirical results.

In this paper, a method to study how the design of indoor spaces and people's movement within them affect disease spread is proposed by integrating computer-aided modeling, multi-agent movement simulation, and airborne viral transmission modeling. Topologicpy spatial design and analysis software is used to model indoor environments, connect spaces, and construct a navigation graph. Pathways for agents, each with unique characteristics such as walking speed, infection status, and activities, are computed using this graph. Agents follow a schedule of events with specific locations and times. The software calculates "time-to-leave" based on walking speed and event start times, and agents are moved along the shortest path within the navigation graph, accurately considering obstacles, doorways, and walls. Precise distance calculations between agents are enabled by this setup. Viral aerosol concentration is then computed and visualized using a reaction-diffusion equation, and each agent's infection risk is determined with an extension of the Wells-Riley ansatz. Infection risk simulations are improved by this spatio-temporal and topological approach, incorporating realistic human behavior and spatial dynamics. The resulting software is designed as a rapid decision-support tool for policymakers, facility managers, stakeholders, architects, and engineers to mitigate disease spread in existing buildings and inform the design of new ones. The software's effectiveness is demonstrated through a comparative analysis of cellular and open commercial office plan layouts.

What data or environments to use for training to improve downstream performance is a longstanding and very topical question in reinforcement learning. In particular, Unsupervised Environment Design (UED) methods have gained recent attention as their adaptive curricula enable agents to be robust to in- and out-of-distribution tasks. We ask to what extent these methods are themselves robust when applied to a novel setting, closely inspired by a real-world robotics problem. Surprisingly, we find that the state-of-the-art UED methods either do not improve upon the na\"{i}ve baseline of Domain Randomisation (DR), or require substantial hyperparameter tuning to do so. Our analysis shows that this is due to their underlying scoring functions failing to predict intuitive measures of ``learnability'', i.e., in finding the settings that the agent sometimes solves, but not always. Based on this, we instead directly train on levels with high learnability and find that this simple and intuitive approach outperforms UED methods and DR in several binary-outcome environments, including on our domain and the standard UED domain of Minigrid. We further introduce a new adversarial evaluation procedure for directly measuring robustness, closely mirroring the conditional value at risk (CVaR). We open-source all our code and present visualisations of final policies here: https://github.com/amacrutherford/sampling-for-learnability.

We introduce a novel dataset for multi-robot activity recognition (MRAR) using two robotic arms integrating WiFi channel state information (CSI), video, and audio data. This multimodal dataset utilizes signals of opportunity, leveraging existing WiFi infrastructure to provide detailed indoor environmental sensing without additional sensor deployment. Data were collected using two Franka Emika robotic arms, complemented by three cameras, three WiFi sniffers to collect CSI, and three microphones capturing distinct yet complementary audio data streams. The combination of CSI, visual, and auditory data can enhance robustness and accuracy in MRAR. This comprehensive dataset enables a holistic understanding of robotic environments, facilitating advanced autonomous operations that mimic human-like perception and interaction. By repurposing ubiquitous WiFi signals for environmental sensing, this dataset offers significant potential aiming to advance robotic perception and autonomous systems. It provides a valuable resource for developing sophisticated decision-making and adaptive capabilities in dynamic environments.

Robotics can help address the growing worker shortage challenge of the manufacturing industry. As such, machine tending is a task collaborative robots can tackle that can also highly boost productivity. Nevertheless, existing robotics systems deployed in that sector rely on a fixed single-arm setup, whereas mobile robots can provide more flexibility and scalability. In this work, we introduce a multi-agent multi-machine tending learning framework by mobile robots based on Multi-agent Reinforcement Learning (MARL) techniques with the design of a suitable observation and reward. Moreover, an attention-based encoding mechanism is developed and integrated into Multi-agent Proximal Policy Optimization (MAPPO) algorithm to boost its performance for machine tending scenarios. Our model (AB-MAPPO) outperformed MAPPO in this new challenging scenario in terms of task success, safety, and resources utilization. Furthermore, we provided an extensive ablation study to support our various design decisions.

Service robots work in a changing environment habited by exogenous agents like humans. In the service robotics domain, lots of uncertainties result from exogenous actions and inaccurate localisation of objects and the robot itself. This makes the robot task scheduling problem incredibly challenging. In this article, we propose a benchmarking system for systematically assessing the performance of algorithms scheduling robot tasks. The robot environment incorporates a room map, furniture, transportable objects, and moving humans; the system defines interfaces for the algorithms, tasks to be executed, and evaluation methods. The system consists of several tools, easing testing scenario generation for training AI-based scheduling algorithms and statistical testing. For benchmarking purposes, a set of scenarios is chosen, and the performance of several scheduling algorithms is assessed. The system source is published to serve the community for tuning and comparable assessment of robot task scheduling algorithms for service robots.

This thesis investigates three areas targeted at improving the reliability of machine learning; fairness in machine learning, strategic classification, and algorithmic robustness. Each of these domains has special properties or structure that can complicate learning. A theme throughout this thesis is thinking about ways in which a'plain' empirical risk minimization algorithm will be misleading or ineffective because of a mis-match between classical learning theory assumptions and specific properties of some data distribution in the wild. The overarching research goal for these related topics is to provide a crisp mathematical model for each learning scenario that exposes different failure modes and makes trade-offs between important metrics explicit in order to provide algorithmic advice or recommendations to practitioners and expose gaps for future research. By tuning our learning algorithms to be more distribution specific in these scenarios, the resulting learned system will exhibit higher utility and avoid catastrophic failure modes. This research is grounded in the theory of machine learning and is fundamentally mathematical in nature, with empirical support when appropriate. Theory is particularly important in these sensitive domains as it is unclear which poor behavior in deployed systems is a natural or benign consequence of a learning system with the underlying distribution,contrasting with problematic but correctable behavior caused by an error in algorithm design or implementation, how to mitigate these issues, or what a successful outcome even looks like in each problem. Theoretical understanding in each domain can help guide best practices and allow for the design of effective, reliable, and robust systems.

A key problem in agent-based simulation is that integrating qualitative insights from multiple discipline experts is extremely hard. In most simulations, agent capabilities and corresponding behaviour needs to be programmed into the agent. We report on the architecture of a tool that disconnects the programmed functions of the agent, from the acquisition of capability and displayed behaviour. This allows multiple different domain experts to represent qualitative insights, without the need for code to be changed. It also allows a continuous integration (or even change) of qualitative behaviour processes, as more insights are gained. The consequent behaviour observed in the model is both, more faithful to the expert's insight as well as able to be contrasted against other models representing other insights.