Weihao Xuan, Rui Yang, Heli Qi, Qingcheng Zeng, Yunze Xiao, Aosong Feng, Dairui Liu, Yun Xing, Junjue Wang, Fan Gao, Jinghui Lu, Yuang Jiang, Huitao Li, Xin Li, Kunyu Yu, Ruihai Dong, Shangding Gu, Yuekang Li, Xiaofei Xie, Felix Juefei-Xu, Foutse Khomh, Osamu Yoshie, Qingyu Chen, Douglas Teodoro, Nan Liu, Randy Goebel, Lei Ma, Edison Marrese-Taylor, Shijian Lu, Yusuke Iwasawa, Yutaka Matsuo, Irene Li. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Driven by inherent uncertainty and the sim-to-real gap, robust reinforcement learning (RL) seeks to improve resilience against the complexity and variability in agent-environment sequential interactions. Despite the existence of a large number of RL benchmarks, there is a lack of standardized benchmarks for robust RL. Current robust RL policies often focus on a specific type of uncertainty and are evaluated in distinct, one-off environments. In this work, we introduce Robust-Gymnasium, a unified modular benchmark designed for robust RL that supports a wide variety of disruptions across all key RL components—agents' observed state and reward, agents' actions, and the environment. Offering over sixty diverse task environments spanning control and robotics, safe RL, and multi-agent RL, it provides an open-source and user-friendly tool for the community to assess current methods and foster the development of robust RL algorithms. In addition, we benchmark existing standard and robust RL algorithms within this framework, uncovering significant deficiencies in each and offering new insights.

Constrained reinforcement learning (RL) seeks high-performance policies under safety constraints. We focus on an offline setting where the agent learns from a fixed dataset—a common requirement in realistic tasks to prevent unsafe exploration. To address this, we propose Diffusion-Regularized Constrained Offline Reinforcement Learning (DRCORL), which first uses a diffusion model to capture the behavioral policy from offline data and then extracts a simplified policy to enable efficient inference. We further apply gradient manipulation for safety adaptation, balancing the reward objective and constraint satisfaction. This approach leverages high-quality offline data while incorporating safety requirements. Empirical results show that DRCORL achieves reliable safety performance, fast inference, and strong reward outcomes across robot learning tasks. Compared to existing safe offline RL methods, it consistently meets cost limits and performs well with the same hyperparameters, indicating practical applicability in real-world scenarios. We open-source our implementation at https://github.com/JamesJunyuGuo/DRCORL.

In numerous reinforcement learning (RL) problems involving safety-critical systems, a key challenge lies in balancing multiple objectives while simultaneously meeting all stringent safety constraints. To tackle this issue, we propose a primal-based framework that orchestrates policy optimization between multi-objective learning and constraint adherence. Our method employs a novel natural policy gradient manipulation method to optimize multiple RL objectives and overcome conflicting gradients between different objectives, since the simple weighted average gradient direction may not be beneficial for specific objectives due to misaligned gradients of different objectives. When there is a violation of a hard constraint, our algorithm steps in to rectify the policy to minimize this violation. Particularly, We establish theoretical convergence and constraint violation guarantees, and our proposed method also outperforms prior state-of-the-art methods on challenging safe multi-objective RL tasks.

Reinforcement Learning (RL) has demonstrated remarkable success across various domains. Nonetheless, a significant challenge in RL is to ensure safety, particularly when deploying it in safety-critical applications such as robotics and autonomous driving. In this work, we develop a robust and safe RL methodology grounded in manifold space. Initially, we construct a constrained manifold space, taking safety constraints into consideration. We then propose a robust safe RL approach, supported by theoretical analysis, based on the value at risk and conditional value at risk, in order to enhance the robustness of safety. Our methodology is designed to ensure safety within stochastic constraint environments. Following the theoretical analysis, we develop a practical, safe algorithm to search for a robust safe policy on stochastic constraint manifolds (ROSCOM). We evaluate the effectiveness of our approach through circular motion and air-hockey tasks. Our experiments demonstrate that ROSCOM outperforms existing baselines in terms of both reward and safety. Note to Practitioners—Real-world applications often involve inherent uncertainties, noise, and high-dimensional spaces. This complexity accentuates the urgency and challenge of ensuring safety in robot learning, especially when implementing RL in practical environments. To address this critical issue, we build a stochastic constraint manifold to delineate the safety space, thus establishing a rigorous framework for robot learning at each iteration. Compared with state-of-the-art baselines, our method can provide remarkable performance regarding safety and reward performance. For example, in an air hockey robot learning task, our method has demonstrated a remarkable 50% enhancement in safety performance compared to the ATACOM framework, while concurrently exhibiting superior reward performance. Moreover, in contrast to traditional algorithms, including CPO, PCPO, our method has achieved a 99% improvement in safety performance, coupled with significantly superior reward performance. These empirical insights render our approach not only theoretically sound but also practically efficacious, indicating its potential as a useful tool in real robot learning and beyond.

The promotion of large-scale applications of reinforcement learning (RL) requires efficient training computation. While existing parallel RL frameworks encompass a variety of RL algorithms and parallelization techniques, the excessively burdensome communication frameworks hinder the attainment of the hardware's limit for final throughput and training effects on a single desktop. In this article, we propose Spreeze, a lightweight parallel framework for RL that efficiently utilizes a single desktop hardware resource to approach the throughput limit. We asynchronously parallelize the experience sampling, network update, performance evaluation, and visualization operations, and employ multiple efficient data transmission techniques to transfer various types of data between processes. The framework can automatically adjust the parallelization hyperparameters based on the computing ability of the hardware device in order to perform efficient large-batch updates. Based on the characteristics of the “Actor-Critic” RL algorithm, our framework uses dual GPUs to independently update the network of actors and critics in order to further improve throughput. Simulation results show that our framework can achieve up to 15,000 Hz experience sampling and 370,000 Hz network update frame rate using only a personal desktop computer, which is an order of magnitude higher than other mainstream parallel RL frameworks, resulting in a 73% reduction of training time. Our work on fully utilizing the hardware resources of a single desktop computer is fundamental to enabling efficient large-scale distributed RL training.

Ensuring the safety of Reinforcement Learning (RL) is crucial for its deployment in real-world applications. Nevertheless, managing the trade-off between reward and safety during exploration presents a significant challenge. Improving reward performance through policy adjustments may adversely affect safety performance. In this study, we aim to address this conflicting relation by leveraging the theory of gradient manipulation. Initially, we analyze the conflict between reward and safety gradients. Subsequently, we tackle the balance between reward and safety optimization by proposing a soft switching policy optimization method, for which we provide convergence analysis. Based on our theoretical examination, we provide a safe RL framework to overcome the aforementioned challenge, and we develop a Safety-MuJoCo Benchmark to assess the performance of safe RL algorithms. Finally, we evaluate the effectiveness of our method on the Safety-MuJoCo Benchmark and a popular safe benchmark, Omnisafe. Experimental results demonstrate that our algorithms outperform several state-of-the-art baselines in terms of balancing reward and safety optimization.

Safe reinforcement learning (RL) is crucial for deploying RL agents in real-world applications, as it aims to maximize long-term rewards while satisfying safety constraints. However, safe RL often suffers from sample inefficiency, requiring extensive interactions with the environment to learn a safe policy. We propose Efficient Safe Policy Optimization (ESPO), a novel approach that enhances the efficiency of safe RL through sample manipulation. ESPO employs an optimization framework with three modes: maximizing rewards, minimizing costs, and balancing the trade-off between the two. By dynamically adjusting the sampling process based on the observed conflict between reward and safety gradients, ESPO theoretically guarantees convergence, optimization stability, and improved sample complexity bounds. Experiments on the Safety-MuJoCo and Omnisafe benchmarks demonstrate that ESPO significantly outperforms existing primal-based and primal-dual-based baselines in terms of reward maximization and constraint satisfaction. Moreover, ESPO achieves substantial gains in sample efficiency, requiring 25--29\% fewer samples than baselines, and reduces training time by 21--38\%.

Reinforcement Learning (RL) has achieved tremendous success in many complex decision-making tasks. However, safety concerns are raised during deploying RL in real-world applications, leading to a growing demand for safe RL algorithms, such as in autonomous driving and robotics scenarios. While safe control has a long history, the study of safe RL algorithms is still in the early stages. To establish a good foundation for future safe RL research, in this paper, we provide a review of safe RL from the perspectives of methods, theories, and applications. First, we review the progress of safe RL from five dimensions and come up with five crucial problems for safe RL being deployed in real-world applications, coined as "2H3W". Second, we analyze the algorithm and theory progress from the perspectives of answering the "2H3W" problems. Particularly, the sample complexity of safe RL algorithms is reviewed and discussed, followed by an introduction to the applications and benchmarks of safe RL algorithms. Finally, we open the discussion of the challenging problems in safe RL, hoping to inspire future research on this thread. To advance the study of safe RL algorithms, we release an open-sourced repository containing major safe RL algorithms at the link.

Industries, such as manufacturing, are accelerating their embrace of the metaverse to achieve higher productivity, especially in complex industrial scheduling. In view of the growing parking challenges in large cities, high-density vehicle spatial scheduling is one of the potential solutions. Stack-based parking lots utilize parking robots to densely park vehicles in the vertical stacks like container stacking, which greatly reduces the aisle area in the parking lot, but requires complex scheduling algorithms to park and take out the vehicles. The existing high-density parking (HDP) scheduling algorithms are mainly heuristic methods, which only contain simple logic and are difficult to utilize information effectively. We propose a hybrid residual multiexpert (HIRE) reinforcement learning (RL) approach, a method for interactive learning in the digital industrial metaverse, which efficiently solves the HDP batch space scheduling problem. In our proposed framework, each heuristic scheduling method is considered as an expert. The neural network trained by RL assigns the expert strategy according to the current parking lot state. Furthermore, to avoid being limited by heuristic expert performance, the proposed hierarchical network framework also sets up a residual output channel. Experiments show that our proposed algorithm outperforms various advanced heuristic methods and the end-to-end RL method in the number of vehicle maneuvers, and has good robustness to the parking lot size and the estimation accuracy of vehicle exit time. We believe that the proposed HIRE RL method can be effectively and conveniently applied to practical application scenarios, which can be regarded as a key step for RL to enter the practical application stage of the industrial metaverse.

Due to a lack of safety considerations, a wide range of multiagent reinforcement learning (MARL) applications are limited in real-world environments. Thus, ensuring MARL safety is essential and urgent in the domain. However, merely a few studies consider the safe MARL problem, and the investigation of real-world applications using safe MARL algorithms still needs to be improved. To fill this gap, we provide a framework with soft constrained policy optimization, in which we develop practical algorithms to address the problem in a cooperative game setting. First, the problem formulation of safe MARL is introduced. Second, the safe policy optimization of safe MARL algorithms based on soft constrained optimization is analyzed, and we further propose a safe learning framework for safe MARL. The framework can be plugged into MARL algorithms without manually fine-tuning safety bounds. Third, we investigate the sim-to-real problems, and conduct simulation and real-world experiments to evaluate the effectiveness of our algorithms. Finally, the comprehensive experimental results indicate that our method has significant benefits regarding the balance between reward and safety performance and outperforms several strong baselines.

A challenging problem in robotics is how to control multiple robots cooperatively and safely in real-world applications. Yet, developing multi-robot control methods from the perspective of safe multi-agent reinforcement learning (MARL) has merely been studied. To fill this gap, in this study, we investigate safe MARL for multi-robot control on cooperative tasks, in which each individual robot has to not only meet its own safety constraints while maximising their reward, but also consider those of others to guarantee safe team behaviours. Firstly, we formulate the safe MARL problem as a constrained Markov game and employ policy optimisation to solve it theoretically. The proposed algorithm guarantees monotonic improvement in reward and satisfaction of safety constraints at every iteration. Secondly, as approximations to the theoretical solution, we propose two safe multi-agent policy gradient methods: Multi-Agent Constrained Policy Optimisation (MACPO) and MAPPO-Lagrangian . Thirdly, we develop the first three safe MARL benchmarks—Safe Multi-Agent MuJoCo (Safe MAMuJoCo), Safe Multi-Agent Robosuite (Safe MARobosuite) and Safe Multi-Agent Isaac Gym (Safe MAIG) to expand the toolkit of MARL and robot control research communities. Finally, experimental results on the three safe MARL benchmarks indicate that our methods can achieve state-of-the-art performance in the balance between improving reward and satisfying safety constraints compared with strong baselines. Demos and code are available at the link ( https://sites.google.com/view/aij-safe-marl/ ). 2

Policy gradient (PG) methods are popular reinforcement learning (RL) methods where a baseline is often applied to reduce the variance of gradient estimates. In multi-agent RL (MARL), although the PG theorem can be naturally extended, the effectiveness of multi-agent PG (MAPG) methods degrades as the variance of gradient estimates increases rapidly with the number of agents. In this paper, we offer a rigorous analysis of MAPG methods by, firstly, quantifying the contributions of the number of agents and agents' explorations to the variance of MAPG estimators. Based on this analysis, we derive the optimal baseline (OB) that achieves the minimal variance. In comparison to the OB, we measure the excess variance of existing MARL algorithms such as vanilla MAPG and COMA. Considering using deep neural networks, we also propose a surrogate version of OB, which can be seamlessly plugged into any existing PG methods in MARL. On benchmarks of Multi-Agent MuJoCo and StarCraft challenges, our OB technique effectively stabilises training and improves the performance of multi-agent PPO and COMA algorithms by a significant margin. Code is released at \url{https://github.com/morning9393/Optimal-Baseline-for-Multi-agent-Policy-Gradients}.