We study learning multi-task, multi-agent policies for cooperative, temporal objectives, under centralized training, decentralized execution. In this setting, using automata to represent tasks assigned to agents enables breaking down a team-level objective into simpler, smaller sub-tasks. However, existing approaches remain sample-inefficient and are limited to the single-task case, requiring retraining policies for each new task. In this work, we present Automata-Conditioned Cooperative Multi-Agent Reinforcement Learning (ACC-MARL), a framework for learning task-conditioned, decentralized team policies. We identify the main challenges to the feasibility of ACC-MARL, propose solutions, and prove that our approach is optimal. We further show that learned value functions can be used to assign tasks optimally at test time. Experiments demonstrate emergent task-aware, multi-step coordination among agents, such as pressing a button to unlock a door, holding the door, and short-circuiting tasks.

Ensuring safety of learning-enabled robotic manipulation across diverse embodiments and tasks still requires significant manual engineering. Existing approaches typically rely on heuristically designed fallback controllers or complex forward invariance assessments. These methods are often too conservative for task success, too computationally expensive for real-time execution, too heuristic to provide useful safety guarantees, or too engineering-heavy to transfer between setups. In this paper, we propose a universal safeguarding approach, X-Safe, which reasons directly in the robot's configuration space to provide formal probabilistic guarantees for collision avoidance. By operating in the configuration space, our method transfers across embodiments while relying solely on an object-based, quasi-static scene representation and a forward kinematics model of the robotic manipulator. Thus, X-Safe provides useful formal safety guarantees without requiring additional data, or engineering effort for different embodiments or scenes. We demonstrate X-Safe for diverse embodiments and policies, both in simulation and on hardware. We observe less degradation in task performance compared to state-of-the-art safeguarding, no collisions on hardware experiments, and empirically corroborate our formal guarantees.

Can a single LLM-based optimization system match specialized tools across fundamentally different domains? We show that when optimization problems are formulated as improving a text artifact evaluated by a scoring function, a single AI-based optimization system-supporting single-task search, multi-task search with cross-problem transfer, and generalization to unseen inputs-achieves state-of-the-art results across six diverse tasks. Our system discovers agent architectures that nearly triple Gemini Flash's ARC-AGI accuracy (32.5

Code-as-Policies (CaP) has shown that large language models (LLMs) can write code to solve robotics tasks by composing perception, planning, and control primitives. Recent CaP systems, however, rely on multi-turn code-generation loops at test time, which is often infeasible for real-time robot control. We introduce Robotics Harness Optimization (RHO), a novel paradigm in which tool-enabled coding agents, at training time, propose and search for interpretable, neurosymbolic multi-file policy repositories (Repositories-as-Policies) that compose these primitives rather than a single prompt, function, or file. RHO searches with reflective feedback from environment reward and execution rather than teleoperation demonstrations. It generalizes to perturbed pick-and-place settings like LIBERO-PRO, where OpenVLA scores 0.0% and $π_{0.5}$ averages 12.83%. Using the same low-level primitives, RHO reaches a 45.0% success rate, 2.5x higher than the strongest multi-turn agentic system, and 3.5x higher than $π_{0.5}$. On Robosuite, RHO sets a new state-of-the-art of 70.0%, exceeding the prior multi-turn record of 68.29% using single-turn execution with no corrective LLM code edits at deployment. When an LLM is used in the control loop, as on RAI's O3DE benchmark, RHO optimizes the deployed agent's multi-file harness of prompts, tools, and control code, improving held-out success from 23.5% to 44.3% with 20% less wall-clock time and 27% fewer tool calls.

Developing autonomous driving systems for complex traffic environments requires balancing multiple objectives, such as avoiding collisions, obeying traffic rules, and making efficient progress. In many situations, these objectives cannot be satisfied simultaneously, and explicit priority relations naturally arise. Also, driving rules require context, so it is important to formally model the environment scenarios within which such rules apply. Existing benchmarks for evaluating autonomous vehicles lack such combinations of multi-objective prioritized rules and formal environment models. In this work, we introduce ScenicRules, a benchmark for evaluating autonomous driving systems in stochastic environments under prioritized multi-objective specifications. We first formalize a diverse set of objectives to serve as quantitative evaluation metrics. Next, we design a Hierarchical Rulebook framework that encodes multiple objectives and their priority relations in an interpretable and adaptable manner. We then construct a compact yet representative collection of scenarios spanning diverse driving contexts and near-accident situations, formally modeled in the Scenic language. Experimental results show that our formalized objectives and Hierarchical Rulebooks align well with human driving judgments and that our benchmark effectively exposes agent failures with respect to the prioritized objectives. Our benchmark can be accessed at https://github.com/BerkeleyLearnVerify/ScenicRules/.

Solving complex real-world control tasks often takes multiple tries: if we fail at first, we reflect on what went wrong, and change our strategy accordingly to avoid making the same mistake. In robotics, Vision-Language-Action models (VLAs) offer a promising path towards solving complex control tasks, but lack the ability to contextually and dynamically readjust behavior when they fail to accomplish a task. In this work, we introduce Learning from Inference-Time Execution (LITEN), which connects a VLA low-level policy to a high-level VLM that conditions on past experiences by including them in-context, allowing it to learn the affordances and capabilities of the low-level VLA. Our approach iterates between a reasoning phase that generates and executes plans for the low-level VLA, and an assessment phase that reflects on the resulting execution and draws useful conclusions to be included in future reasoning contexts. Unlike similar approaches to self-refinement in non-robotics domains, LITEN must reflect on unstructured real-world robot trajectories (e.g., raw videos), which requires careful treatment during assessment. Our experimental results demonstrate LITEN is able to effectively learn from past experience to generate plans that use high-affordance instructions to accomplish long-horizon tasks.

In certifiable machine learning, AI systems produce not only results but also verifiable certificates that the results can be trusted.

We introduce a novel framework for learning context-aware runtime monitors for AI-based control ensembles. Machine-learning (ML) controllers are increasingly deployed in (autonomous) cyber-physical systems because of their ability to solve complex decision-making tasks. However, their accuracy can degrade sharply in unfamiliar environments, creating significant safety concerns. Traditional ensemble methods aim to improve robustness by averaging or voting across multiple controllers, yet this often dilutes the specialized strengths that individual controllers exhibit in different operating contexts. We argue that, rather than blending controller outputs, a monitoring framework should identify and exploit these contextual strengths. In this paper, we reformulate the design of safe AI-based control ensembles as a contextual monitoring problem. A monitor continuously observes the system's context and selects the controller best suited to the current conditions. To achieve this, we cast monitor learning as a contextual learning task and draw on techniques from contextual multi-armed bandits. Our approach comes with two key benefits: (1) theoretical safety guarantees during controller selection, and (2) improved utilization of controller diversity. We validate our framework in two simulated autonomous driving scenarios, demonstrating significant improvements in both safety and performance compared to non-contextual baselines.

We introduce a novel framework for learning context-aware runtime monitors for AI-based control ensembles. Machine-learning (ML)-based controllers are increasingly deployed in (autonomous) cyber–physical systems due to their ability to solve complex decision-making tasks. However, their accuracy can degrade in unfamiliar environments, creating significant safety concerns. Traditional ensemble methods aim to improve robustness by averaging or voting across multiple controllers, but this often dilutes the specialized strengths of individual controllers in different operating contexts. We argue that a monitoring framework should identify and exploit these contextual strengths. We reformulate the design of safe AIbased control ensembles as a contextual monitoring problem. A monitor continuously observes the system’s context and selects the controller best suited to the current conditions. We cast monitor learning as a contextual learning task and draw on techniques from contextual multi-armed bandits. Our approach brings two key benefits: (1) theoretical safety guarantees during controller selection, and (2) improved utilization of controller diversity. We validate our framework in two simulated autonomous driving scenarios, showing significant improvements in safety and performance, specifically over non-contextual baselines.

Vision Language Models (VLMs) achieve strong performance on many vision-language tasks but often struggle with spatial reasoning—a prerequisite for many applications. Empirically, we find that a dataset produced by a current training data generation pipeline has a 57.6% human validation rate. These rates stem from current limitations: single-image 3D reconstruction introduces cascading modeling errors and requires wide answer tolerances, while caption-based methods require hyper-detailed annotations and suffer from generative hallucinations. We present GRAID, built on the key insight that qualitative spatial relationships can be reliably determined from 2D geometric primitives alone. By operating exclusively on 2D bounding boxes from standard object detectors, GRAID avoids both 3D reconstruction errors and generative hallucinations, resulting in datasets that are of higher quality than existing tools that produce similar datasets as validated by human evaluations. We apply our framework to the BDD100k, NuImages, and Waymo datasets, generating over 8.5 million high-quality VQA pairs creating questions spanning spatial relations, counting, ranking, and size comparisons. We evaluate one of the datasets and find it achieves 91.16% human-validated accuracy—compared to 57.6% on a dataset generated by recent work.

Simulation-based testing has become a crucial complement to road testing for ensuring the safety of cyber-physical systems (CPS). As a result, significant research efforts have been directed toward identifying failure scenarios within simulation environments. However, a critical question remains: are the AV failure scenarios discovered in simulation relevant to real-world systems-specifically, are they reproducible on actual systems? The sim-to-real gap caused by differences between simulated and real sensor data means that failure scenarios identified in simulation might either be artifacts of synthetic sensor data or actual issues that also occur with real sensor data. To address this, an effective approach to validating simulated failure scenarios is to locate occurrences of these scenarios within real-world datasets and verify whether the failure persists on the datasets. To this end, we introduce a formal definition of how labeled time series sensor data can match an abstract scenario, represented as a scenario program using the Scenic probabilistic programming language. We present a querying algorithm that, given a scenario program and a labeled dataset, identifies the subset of data that matches the specified scenario. Our experiment shows that our algorithm is more accurate and orders of magnitude faster in querying scenarios than the state-of-the-art commercial vision large language models, and can scale with the duration of queried time series data.

Safety-critical autonomous systems must reliably predict unsafe behavior to take timely corrective actions. Safety properties are often defined over variables that are not directly observable at runtime, making prediction and detection of violations hard. We present a new approach for learning interpretable monitors characterized by concise Signal Temporal Logic (STL) formulas that can predict safety property violations from the observable sensor data. We train these monitors from synthetic, possibly highly unbalanced data generated in a simulation environment. Our specification mining procedure combines a grammar-based method and two novel ensemble techniques. Our approach outperforms the existing solutions by enhancing accuracy and explainability, as demonstrated in two autonomous driving case studies.

Contract-based design is a method to facilitate modular design of systems. While there has been substantial progress on the theory of contracts, there has been less progress on practical algorithms for the algebraic operations in the theory. In this article, we present (1) principles to implement a contract-based design tool at scale and (2) Pacti, a tool that can efficiently compute these operations. We illustrate the use of Pacti in a variety of case studies.

Automata-conditioned reinforcement learning (RL) has given promising results for learning multi-task policies capable of performing temporally extended objectives given at runtime, done by pretraining and freezing automata embeddings prior to training the downstream policy. However, no theoretical guarantees were given. This work provides a theoretical framework for the automata-conditioned RL problem and shows that it is probably approximately correct learnable. We then present a technique for learning provably correct automata embeddings, guaranteeing optimal multi-task policy learning. Our experimental evaluation confirms these theoretical results.

Digital health interventions are increasingly used in physical and occupational therapy to deliver home exercise programs via sensor equipped devices such as smartphones, enabling remote monitoring of adherence and performance. However, digital interventions are typically programmed as software before clinical encounters as libraries of parametrized exercise modules targeting broad patient populations. At the point of care, clinicians can only select modules and adjust a narrow set of parameters like repetitions, so patient specific needs that emerge during encounters, such as distinct movement limitations, and home environments, are rarely reflected in the software. We evaluated a digital intervention paradigm that uses large language models (LLMs) to translate clinicians' exercise prescriptions into intervention software. In a prospective single arm feasibility study with 20 licensed physical and occupational therapists and a standardized patient, clinicians created 40 individualized upper extremity programs (398 instructions) that were automatically translated into executable software. Our results show a 45

Clients rely on database systems to be correct, which requires the system not only to implement transactions’ semantics correctly but also to provide isolation guarantees for the transactions. This paper presents a client-centric technique for checking both semantic correctness and isolation-level guarantees for black-box database systems based on observations collected from running transactions on these systems. Our technique verifies observational correctness with respect to a given set of transactions and observations for them, which holds iff there exists a possible correct execution of the transactions under a given isolation level that could result in these observations. Our technique relies on novel symbolic encodings of (1) the semantic correctness of database transactions in the presence of weak isolation and (2) isolation-level guarantees. These are used by the checker to query a Satisfiability Modulo Theories solver. We applied our tool Troubadour to verify observational correctness of several database systems, including PostgreSQL and an industrial system under development, in which the tool helped detect two new bugs. We also demonstrate that Troubadour is able to find known semantic correctness bugs and detect isolation-related anomalies.

To design performant, expressive, and reliable cyber-physical systems (CPSs), researchers extensively perform quasi-static scheduling for concurrent models of computation (MoCs) on multi-core hardware. However, these quasi-static scheduling approaches are developed independently for their corresponding MoCs, despite commonality in the approaches. To help generalize the use of quasi-static scheduling to new and emerging MoCs, this article proposes a unified approach for a class of deterministic timed concurrent models (DTCMs), including prominent models such as synchronous dataflow (SDF), Boolean-controlled dataflow (BDF), scenario-aware dataflow (SADF), and Logical Execution Time (LET). In contrast to scheduling techniques tailored exclusively to specific MoCs, our unified approach leverages a common intermediate formalism called state space finite automata (SSFA), bridging the gap between high-level MoCs and executable schedules. Once identified as DTCMs, new MoCs can directly adopt SSFA-based scheduling, significantly easing adoption. We show that quasi-static schedules facilitated by SSFA are provably free from timing anomalies and enable straightforward worst-case makespan analysis. We demonstrate the approach using the reactor model—an emerging discrete-event MoC—programmed using the Lingua Franca (LF) language. Experiments show that quasi-statically scheduled LF programs exhibit lower runtime overhead compared to the dynamically scheduled LF programs, and that the analyzable worst-case makespans enable compile-time deadline checking.