Reinforcement learning (RL) has gained traction for its success in solving complex tasks for robotic applications. However, its deployment on physical robots remains challenging due to safety risks and the comparatively high costs of training. To avoid these problems, RL agents are often trained on simulators, which introduces a new problem related to the gap between simulation and reality. This paper presents an RL pipeline designed to help reduce the reality gap and facilitate developing and deploying RL policies for real-world robotic systems. The pipeline organizes the RL training process into an initial step for system identification and three training stages: core simulation training, high-fidelity simulation, and real-world deployment, each adding levels of realism to reduce the sim-to-real gap. Each training stage takes an input policy, improves it, and either passes the improved policy to the next stage or loops it back for further improvement. This iterative process continues until the policy achieves the desired performance. The pipeline's effectiveness is shown through a case study with the Boston Dynamics Spot mobile robot used in a surveillance application. The case study presents the steps taken at each pipeline stage to obtain an RL agent to control the robot's position and orientation.

We study infinite-horizon Discounted Markov Decision Processes (DMDPs) under a generative model. Motivated by the Algorithm with Advice framework Mitzenmacher and Vassilvitskii 2022, we propose a novel framework to investigate how a prediction on the transition matrix can enhance the sample efficiency in solving DMDPs and improve sample complexity bounds. We focus on the DMDPs with $N$ state-action pairs and discounted factor $\gamma$. Firstly, we provide an impossibility result that, without prior knowledge of the prediction accuracy, no sampling policy can compute an $\epsilon$-optimal policy with a sample complexity bound better than $\tilde{O}((1-\gamma)^{-3} N\epsilon^{-2})$, which matches the state-of-the-art minimax sample complexity bound with no prediction. In complement, we propose an algorithm based on minimax optimization techniques that leverages the prediction on the transition matrix. Our algorithm achieves a sample complexity bound depending on the prediction error, and the bound is uniformly better than $\tilde{O}((1-\gamma)^{-4} N \epsilon^{-2})$, the previous best result derived from convex optimization methods. These theoretical findings are further supported by our numerical experiments.

Embodied multimodal large models (EMLMs) have gained significant attention in recent years due to their potential to bridge the gap between perception, cognition, and action in complex, real-world environments. This comprehensive review explores the development of such models, including Large Language Models (LLMs), Large Vision Models (LVMs), and other models, while also examining other emerging architectures. We discuss the evolution of EMLMs, with a focus on embodied perception, navigation, interaction, and simulation. Furthermore, the review provides a detailed analysis of the datasets used for training and evaluating these models, highlighting the importance of diverse, high-quality data for effective learning. The paper also identifies key challenges faced by EMLMs, including issues of scalability, generalization, and real-time decision-making. Finally, we outline future directions, emphasizing the integration of multimodal sensing, reasoning, and action to advance the development of increasingly autonomous systems. By providing an in-depth analysis of state-of-the-art methods and identifying critical gaps, this paper aims to inspire future advancements in EMLMs and their applications across diverse domains.

Scaling up the model size and computation has brought consistent performance improvements in supervised learning. However, this lesson often fails to apply to reinforcement learning (RL) because training the model on non-stationary data easily leads to overfitting and unstable optimization. In response, we introduce SimbaV2, a novel RL architecture designed to stabilize optimization by (i) constraining the growth of weight and feature norm by hyperspherical normalization; and (ii) using a distributional value estimation with reward scaling to maintain stable gradients under varying reward magnitudes. Using the soft actor-critic as a base algorithm, SimbaV2 scales up effectively with larger models and greater compute, achieving state-of-the-art performance on 57 continuous control tasks across 4 domains. The code is available at https://dojeon-ai.github.io/SimbaV2.

Reinforcement learning (RL) has shown impressive results in sequential decision-making tasks. Meanwhile, Large Language Models (LLMs) and Vision-Language Models (VLMs) have emerged, exhibiting impressive capabilities in multimodal understanding and reasoning. These advances have led to a surge of research integrating LLMs and VLMs into RL. In this survey, we review representative works in which LLMs and VLMs are used to overcome key challenges in RL, such as lack of prior knowledge, long-horizon planning, and reward design. We present a taxonomy that categorizes these LLM/VLM-assisted RL approaches into three roles: agent, planner, and reward. We conclude by exploring open problems, including grounding, bias mitigation, improved representations, and action advice. By consolidating existing research and identifying future directions, this survey establishes a framework for integrating LLMs and VLMs into RL, advancing approaches that unify natural language and visual understanding with sequential decision-making.

This paper investigates the instability of Q-learning in continuous environments, a challenge frequently encountered by practitioners. Traditionally, this instability is attributed to bootstrapping and regression model errors. Using a representative reinforcement learning benchmark, we systematically examine the effects of bootstrapping and model inaccuracies by incrementally eliminating these potential error sources. Our findings reveal that even in relatively simple benchmarks, the fundamental task of Q-learning - iteratively learning a Q-function from policy-specific target values - can be inherently ill-posed and prone to failure. These insights cast doubt on the reliability of Q-learning as a universal solution for reinforcement learning problems.

--Cooperative intelligence (CI) is expected to become an integral element in next-generation networks because it can aggregate the capabilities and intelligence of multiple devices. Multi-agent reinforcement learning (MARL) is a popular approach for achieving CI in communication problems by enabling effective collaboration among agents to address sequential problems. However, ensuring privacy protection for MARL is a challenging task because of the presence of heterogeneous agents that learn interdependently via sharing information. Implementing privacy protection techniques such as data encryption and federated learning to MARL introduces the notable overheads (e.g., computation and bandwidth). T o overcome these challenges, we propose PP-MARL, an efficient privacy-preserving learning scheme for MARL. PP-MARL leverages homomorphic encryption (HE) and differential privacy (DP) to protect privacy, while introducing split learning to decrease overheads via reducing the volume of shared messages, and then improve efficiency. We apply and evaluate PP-MARL in two communication-related use cases. Simulation results reveal that PP-MARL can achieve efficient and reliable collaboration with 1.1-6 times better privacy protection and lower overheads (e.g., 84-91% reduction in bandwidth) than state-of-the-art approaches. Cooperative intelligence (CI) [1], [2] is expected to facilitate next-generation networks by establishing collaboration among various communication-related intelligent equipment. Multi-agent reinforcement learning (MARL) is a popular approach for achieving CI in addressing sequential problems in communication [3], such as adaptive routing and resource allocation [4].

As artificial intelligence scales, the concepts of alignment, agency, and autonomy have become central to AI safety, governance, and control. However, even in human contexts, these terms lack universal definitions, varying across disciplines such as philosophy, psychology, law, computer science, mathematics, and political science. This inconsistency complicates their application to AI, where differing interpretations lead to conflicting approaches in system design and regulation. This paper traces the historical, philosophical, and technical evolution of these concepts, emphasizing how their definitions influence AI development, deployment, and oversight. We argue that the urgency surrounding AI alignment and autonomy stems not only from technical advancements but also from the increasing deployment of AI in high-stakes decision making. Using Agentic AI as a case study, we examine the emergent properties of machine agency and autonomy, highlighting the risks of misalignment in real-world systems. Through an analysis of automation failures (Tesla Autopilot, Boeing 737 MAX), multi-agent coordination (Metas CICERO), and evolving AI architectures (DeepMinds AlphaZero, OpenAIs AutoGPT), we assess the governance and safety challenges posed by frontier AI.

Ensuring safety in autonomous driving systems remains a critical challenge, particularly in handling rare but potentially catastrophic safety-critical scenarios. While existing research has explored generating safety-critical scenarios for autonomous vehicle (AV) testing, there is limited work on effectively incorporating these scenarios into policy learning to enhance safety. Furthermore, developing training curricula that adapt to an AV's evolving behavioral patterns and performance bottlenecks remains largely unexplored. To address these challenges, we propose CurricuVLM, a novel framework that leverages Vision-Language Models (VLMs) to enable personalized curriculum learning for autonomous driving agents. Our approach uniquely exploits VLMs' multimodal understanding capabilities to analyze agent behavior, identify performance weaknesses, and dynamically generate tailored training scenarios for curriculum adaptation. Through comprehensive analysis of unsafe driving situations with narrative descriptions, CurricuVLM performs in-depth reasoning to evaluate the AV's capabilities and identify critical behavioral patterns. The framework then synthesizes customized training scenarios targeting these identified limitations, enabling effective and personalized curriculum learning. Extensive experiments on the Waymo Open Motion Dataset show that CurricuVLM outperforms state-of-the-art baselines across both regular and safety-critical scenarios, achieving superior performance in terms of navigation success, driving efficiency, and safety metrics. Further analysis reveals that CurricuVLM serves as a general approach that can be integrated with various RL algorithms to enhance autonomous driving systems. The code and demo video are available at: https://zihaosheng.github.io/CurricuVLM/.

A long-standing goal in AI is to build agents that can solve a variety of tasks across different environments, including previously unseen ones. Two dominant approaches tackle this challenge: (i) reinforcement learning (RL), which learns policies through trial and error, and (ii) optimal control, which plans actions using a learned or known dynamics model. However, their relative strengths and weaknesses remain underexplored in the setting where agents must learn from offline trajectories without reward annotations. In this work, we systematically analyze the performance of different RL and control-based methods under datasets of varying quality. On the RL side, we consider goal-conditioned and zero-shot approaches. On the control side, we train a latent dynamics model using the Joint Embedding Predictive Architecture (JEPA) and use it for planning. We study how dataset properties-such as data diversity, trajectory quality, and environment variability-affect the performance of these approaches. Our results show that model-free RL excels when abundant, high-quality data is available, while model-based planning excels in generalization to novel environment layouts, trajectory stitching, and data-efficiency. Notably, planning with a latent dynamics model emerges as a promising approach for zero-shot generalization from suboptimal data.