Flow-matching policies have emerged as a powerful paradigm for generalist robotics. These models are trained to imitate an action chunk, conditioned on sensor observations and textual instructions. Often, training demonstrations are generated by a suboptimal policy, such as a human operator. This work explores training flow-matching policies via reinforcement learning to surpass the original demonstration policy performance. We particularly note minimum-time control as a key application and present a simple scheme for variable-horizon flow-matching planning. We then introduce two families of approaches: a simple Reward-Weighted Flow Matching (RWFM) scheme and a Group Relative Policy Optimization (GRPO) approach with a learned reward surrogate. Our policies are trained on an illustrative suite of simulated unicycle dynamics tasks, and we show that both approaches dramatically improve upon the suboptimal demonstrator performance, with the GRPO approach in particular generally incurring between $50\%$ and $85\%$ less cost than a naive Imitation Learning Flow Matching (ILFM) approach.

Large language models (LLMs), despite possessing latent safety understanding from their vast pretraining data, remain vulnerable to generating harmful content and exhibit issues such as over-refusal and utility degradation after safety alignment. Current safety alignment methods often result in superficial refusal shortcuts or rely on intensive supervision for reasoning-based approaches, failing to fully leverage the model's intrinsic safety self-awareness. We propose \textbf{AlphaAlign}, a simple yet effective pure reinforcement learning (RL) framework with verifiable safety reward designed to incentivize this latent safety awareness through proactive safety reasoning.} AlphaAlign employs a dual-reward system: a verifiable safety reward encourages correctly formatted and explicitly justified refusals for harmful queries while penalizing over-refusals, and a normalized helpfulness reward guides high-quality responses to benign inputs. This allows the model to develop proactive safety reasoning capabilities without depending on supervised safety-specific reasoning data. AlphaAlign demonstrates three key advantages: (1) Simplicity and efficiency, requiring only binary prompt safety labels and minimal RL steps for substantial improvements. (2) Breaking the safety-utility trade-off, by enhancing refusal of harmful content and reducing over-refusals, while simultaneously maintaining or even improving general task performance and robustness to unseen jailbreaks. (3) Deep alignment, fostering proactive safety reasoning that generates explicit safety rationales rather than relying on shallow refusal patterns.

This research introduces the Theory of Partial Symmetry Enforced Attention Decomposition (PSEAD), a new and rigorous group-theoretic framework designed to seamlessly integrate local symmetry awareness into the core architecture of self-attention mechanisms within Transformer models. We formalize the concept of local permutation subgroup actions on windows of biological data, proving that under such actions, the attention mechanism naturally decomposes into a direct sum of orthogonal irreducible components. Critically, these components are intrinsically aligned with the irreducible representations of the acting permutation subgroup, thereby providing a powerful mathematical basis for disentangling symmetric and asymmetric features. We show that PSEAD offers substantial advantages. These include enhanced generalization capabilities to novel biological motifs exhibiting similar partial symmetries, unprecedented interpretability by allowing direct visualization and analysis of attention contributions from different symmetry channels, and significant computational efficiency gains by focusing representational capacity on relevant symmetric subspaces. Beyond static data analysis, we extend PSEAD's applicability to dynamic biological processes within reinforcement learning paradigms, showcasing its potential to accelerate the discovery and optimization of biologically meaningful policies in complex environments like protein folding and drug discovery. This work lays the groundwork for a new generation of biologically informed, symmetry-aware artificial intelligence models.

Autonomous driving demands reliable and efficient solutions to closely related problems such as decision-making and motion planning. In this work, decision-making refers specifically to highway lane selection, while motion planning involves generating control commands (such as speed and steering) to reach the chosen lane. In the context of Connected Autonomous Vehicles (CAVs), achieving both flexible and safe lane selection alongside precise trajectory execution remains a significant challenge. This paper proposes a framework called Cohesive Decision-Guided Motion Planning (CDGMP), which tightly integrates decision-making and motion planning using a Mixture of Experts (MoE) inspired architecture combined with multi-policy reinforcement learning. By coordinating multiple specialized sub-networks through a gating mechanism, the method decomposes the complex driving task into modular components. Each sub-network focuses on a specific aspect of driving, improving efficiency by activating only the most relevant modules during inference. This design also enhances safety through modular specialization. CDGMP improves the adaptability and robustness of CAVs across diverse traffic scenarios, offering a scalable solution to real-world autonomy challenges. The architectural principles behind CDGMP, especially the use of MoE, also provide a strong foundation for other high-dimensional decision and control tasks. Simulation results (available at https://youtu.be/_-4OXNHV0UY) demonstrate reliable performance in both lane selection and motion planning.

Language model (LM) agents have gained significant attention for their ability to autonomously complete tasks through interactions with environments, tools, and APIs. LM agents are primarily built with prompt engineering or supervised finetuning. At the same time, reinforcement learning (RL) has been explored to enhance LM's capabilities, such as reasoning and factuality. However, the combination of the LM agents and reinforcement learning (Agent-RL) remains underexplored and lacks systematic study. To this end, we built AgentFly, a scalable and extensible Agent-RL framework designed to empower LM agents with a variety of RL algorithms. Our framework supports multi-turn interactions by adapting traditional RL methods with token-level masking. It features a decorator-based interface for defining tools and reward functions, enabling seamless extension and ease of use. To support high-throughput training, we implement asynchronous execution of tool calls and reward computations, and design a centralized resource management system for scalable environment coordination. We also provide a suite of prebuilt tools and environments, demonstrating the framework's effectiveness through successful agent training across multiple tasks.

This study addresses the problem of occluded grasping, where primary grasp configurations of an object are not available due to occlusion with environment. Simple parallel grippers often struggle with such tasks due to limited dexterity and actuation constraints. Prior works have explored object pose reorientation such as pivoting by utilizing extrinsic contacts between an object and an environment feature like a wall, to make the object graspable. However, such works often assume the presence of a short wall, and this assumption may not always hold in real-world scenarios. If the wall available for interaction is too large or too tall, the robot may still fail to grasp the object even after pivoting, and the robot must combine different types of actions to grasp. To address this, we propose a hierarchical reinforcement learning (RL) framework. We use Q-learning to train a high-level policy that selects the type of action expected to yield the highest reward. The selected low-level skill then samples a specific robot action in continuous space. To guide the robot to an appropriate location for executing the selected action, we adopt a Conditional Variational Autoencoder (CVAE). We condition the CVAE on the object point cloud and the skill ID, enabling it to infer a suitable location based on the object geometry and the selected skill. To promote generalization, we apply domain randomization during the training of low-level skills. The RL policy is trained entirely in simulation with a box-like object and deployed to six objects in real world. We conduct experiments to evaluate our method and demonstrate both its generalizability and robust sim-to-real transfer performance with promising success rates.

Popular methods in cooperative Multi-Agent Reinforcement Learning with partially observable environments typically allow agents to act independently during execution, which may limit the coordinated effect of the trained policies. However, by sharing information such as known or suspected ongoing threats, effective communication can lead to improved decision-making in the cyber battle space. We propose a game design where defender agents learn to communicate and defend against imminent cyber threats by playing training games in the Cyber Operations Research Gym, using the Differentiable Inter Agent Learning algorithm adapted to the cyber operational environment. The tactical policies learned by these autonomous agents are akin to those of human experts during incident responses to avert cyber threats. In addition, the agents simultaneously learn minimal cost communication messages while learning their defence tactical policies.

We consider the maximum causal entropy inverse reinforcement learning problem for infinite-horizon stationary mean-field games, in which we model the unknown reward function within a reproducing kernel Hilbert space. This allows the inference of rich and potentially nonlinear reward structures directly from expert demonstrations, in contrast to most existing inverse reinforcement learning approaches for mean-field games that typically restrict the reward function to a linear combination of a fixed finite set of basis functions. We also focus on the infinite-horizon cost structure, whereas prior studies primarily rely on finite-horizon formulations. We introduce a Lagrangian relaxation to this maximum causal entropy inverse reinforcement learning problem that enables us to reformulate it as an unconstrained log-likelihood maximization problem, and obtain a solution \lk{via} a gradient ascent algorithm. To illustrate the theoretical consistency of the algorithm, we establish the smoothness of the log-likelihood objective by proving the Fréchet differentiability of the related soft Bellman operators with respect to the parameters in the reproducing kernel Hilbert space. We demonstrate the effectiveness of our method on a mean-field traffic routing game, where it accurately recovers expert behavior.

Abstract--We investigate a Federated Reinforcement Learning with Environment Heterogeneity (FRL-EH) framework, where local environments exhibit statistical heterogeneity . Within this framework, agents collaboratively learn a global policy by aggregating their collective experiences while preserving the privacy of their local trajectories. T o better reflect real-world scenarios, we introduce a robust FRL-EH framework by presenting a novel global objective function. This function is specifically designed to optimize a global policy that ensures robust performance across heterogeneous local environments and their plausible perturbations. We propose a tabular FRL algorithm named FedRQ and theoretically prove its asymptotic convergence to an optimal policy for the global objective function. Furthermore, we extend FedRQ to environments with continuous state space through the use of expectile loss, addressing the key challenge of minimizing a value function over a continuous subset of the state space. Reinforcement Learning (RL) has demonstrated remarkable efficacy in tackling complex challenges across various domains, including gaming, robotics, intelligent networks, manufacturing, and finance [1]-[3]. However, the practical implementation of RL algorithms often encounters persistent obstacles, particularly the scarcity of training samples, especially in large action and state spaces.

Unmanned aerial vehicles (UAVs) equipped with integrated sensing and communication (ISAC) capabilities are envisioned to play a pivotal role in future wireless networks due to their enhanced flexibility and efficiency. However, jointly optimizing UAV trajectory planning, multi-user communication, and target sensing under stringent resource constraints and time-critical conditions remains a significant challenge. To address this, we propose an Age of Information (AoI)-centric UAV-ISAC system that simultaneously performs target sensing and serves multiple ground users, emphasizing information freshness as the core performance metric. We formulate a long-term average AoI minimization problem that jointly optimizes the UAV's flight trajectory and beamforming. To tackle the high-dimensional, non-convexity of this problem, we develop a deep reinforcement learning (DRL)-based algorithm capable of providing real-time decisions on UAV movement and beamforming for both radar sensing and multi-user communication. Specifically, a Kalman filter is employed for accurate target state prediction, regularized zero-forcing is utilized to mitigate inter-user interference, and the Soft Actor-Critic algorithm is applied for training the DRL agent on continuous actions. The proposed framework adaptively balances the trade-offs between sensing accuracy and communication quality. Extensive simulation results demonstrate that our proposed method consistently achieves lower average AoI compared to baseline approaches.