Existing approaches to language model alignment often treat safety as a tradeoff against helpfulness, which can lead to unacceptable responses in sensitive domains. To ensure reliable performance in such settings, we propose High-Confidence Safe Reinforcement Learning from Human Feedback (HC-RLHF), a method that provides high-confidence safety guarantees while maximizing helpfulness. Similar to previous methods, HC-RLHF explicitly decouples human preferences into helpfulness and harmlessness (safety), which are learned by training a reward model and a cost model, respectively. It then employs a two-step process to find safe solutions. In the first step, it optimizes the reward function under an intentionally pessimistic version of the cost constraint. In the second step, the trained model undergoes a safety test to verify whether its performance stays within an upper-confidence bound of the actual cost constraint. We provide a theoretical analysis of HC-RLHF, including proof that it will not return an unsafe solution with a probability greater than a user-specified threshold. For our empirical analysis, we apply HC-RLHF to align three different language models (Qwen2-1.5B, Qwen2.5-3B, and LLaMa3.2-3B) with human preferences. Our results demonstrate that HC-RLHF produces safe models with high probability and can improve harmlessness and helpfulness compared to previous methods.
Deploying autonomous robots in safety-critical applications requires safety guarantees. Provably safe reinforcement learning is an active field of research which aims to provide such guarantees using safeguards. These safeguards should be integrated during training to prevent a large sim-to-real gap. While there are several approaches for safeguarding sampling-based reinforcement learning, analytic gradient-based reinforcement learning often achieves superior performance and sample efficiency. However, there is no safeguarding approach for this learning paradigm yet. Our work addresses this gap by developing the first effective safeguard for analytic gradient-based reinforcement learning. We analyse existing, differentiable safeguards, adapt them through modified mappings and gradient formulations, and integrate them with a state-of-the-art learning algorithm and a differentiable simulation. We evaluate how different safeguards affect policy optimisation using numerical experiments on two classical control tasks. The results demonstrate safeguarded training without compromising performance.
A longstanding goal in safe reinforcement learning (RL) is a method to ensure the safety of a policy throughout the entire process, from learning to operation. However, existing safe RL paradigms inherently struggle to achieve this objective. We propose a method, called Provably Lifetime Safe RL (PLS), that integrates offline safe RL with safe policy deployment to address this challenge. Our proposed method learns a policy offline using return-conditioned supervised learning and then deploys the resulting policy while cautiously optimizing a limited set of parameters, known as target returns, using Gaussian processes (GPs). Theoretically, we justify the use of GPs by analyzing the mathematical relationship between target and actual returns. We then prove that PLS finds near-optimal target returns while guaranteeing safety with high probability. Empirically, we demonstrate that PLS outperforms baselines both in safety and reward performance, thereby achieving the longstanding goal to obtain high rewards while ensuring the safety of a policy throughout the lifetime from learning to operation.
When applying offline reinforcement learning (RL) in healthcare scenarios, the out-of-distribution (OOD) issues pose significant risks, as inappropriate generalization beyond clinical expertise can result in potentially harmful recommendations. While existing methods like conservative Q-learning (CQL) attempt to address the OOD issue, their effectiveness is limited by only constraining action selection by suppressing uncertain actions. This action-only regularization imitates clinician actions that prioritize short-term rewards, but it fails to regulate downstream state trajectories, thereby limiting the discovery of improved long-term treatment strategies. To safely improve policy beyond clinician recommendations while ensuring that state-action trajectories remain in-distribution, we propose \textit{Offline Guarded Safe Reinforcement Learning} ($\mathsf{OGSRL}$), a theoretically grounded model-based offline RL framework. $\mathsf{OGSRL}$ introduces a novel dual constraint mechanism for improving policy with reliability and safety. First, the OOD guardian is established to specify clinically validated regions for safe policy exploration. By constraining optimization within these regions, it enables the reliable exploration of treatment strategies that outperform clinician behavior by leveraging the full patient state history, without drifting into unsupported state-action trajectories. Second, we introduce a safety cost constraint that encodes medical knowledge about physiological safety boundaries, providing domain-specific safeguards even in areas where training data might contain potentially unsafe interventions. Notably, we provide theoretical guarantees on safety and near-optimality: policies that satisfy these constraints remain in safe and reliable regions and achieve performance close to the best possible policy supported by the data.
This paper presents a risk-aware safe reinforcement learning (RL) control design for stochastic discrete-time linear systems. Rather than using a safety certifier to myopically intervene with the RL controller, a risk-informed safe controller is also learned besides the RL controller, and the RL and safe controllers are combined together. Several advantages come along with this approach: 1) High-confidence safety can be certified without relying on a high-fidelity system model and using limited data available, 2) Myopic interventions and convergence to an undesired equilibrium can be avoided by deciding on the contribution of two stabilizing controllers, and 3) highly efficient and computationally tractable solutions can be provided by optimizing over a scalar decision variable and linear programming polyhedral sets. To learn safe controllers with a large invariant set, piecewise affine controllers are learned instead of linear controllers. To this end, the closed-loop system is first represented using collected data, a decision variable, and noise. The effect of the decision variable on the variance of the safe violation of the closed-loop system is formalized. The decision variable is then designed such that the probability of safety violation for the learned closed-loop system is minimized. It is shown that this control-oriented approach reduces the data requirements and can also reduce the variance of safety violations. Finally, to integrate the safe and RL controllers, a new data-driven interpolation technique is introduced. This method aims to maintain the RL agent's optimal implementation while ensuring its safety within environments characterized by noise. The study concludes with a simulation example that serves to validate the theoretical results.
Offline safe reinforcement learning(OSRL) derives constraint-satisfying policies from pre-collected datasets, offers a promising avenue for deploying RL in safety-critical real-world domains such as robotics. However, the majority of existing approaches emphasize only short-term safety, neglecting long-horizon considerations. Consequently, they may violate safety constraints and fail to ensure sustained protection during online deployment. Moreover, the learned policies often struggle to handle states and actions that are not present or out-of-distribution(OOD) from the offline dataset, and exhibit limited sample efficiency. To address these challenges, we propose a novel framework Feasibility-Aware offline Safe Reinforcement Learning with CVAE-based Pessimism (FASP). First, we employ Hamilton-Jacobi (H-J) reachability analysis to generate reliable safety labels, which serve as supervisory signals for training both a conditional variational autoencoder (CVAE) and a safety classifier. This approach not only ensures high sampling efficiency but also provides rigorous long-horizon safety guarantees. Furthermore, we utilize pessimistic estimation methods to estimate the Q-value of reward and cost, which mitigates the extrapolation errors induces by OOD actions, and penalize unsafe actions to enabled the agent to proactively avoid high-risk behaviors. Moreover, we theoretically prove the validity of this pessimistic estimation. Extensive experiments on DSRL benchmarks demonstrate that FASP algorithm achieves competitive performance across multiple experimental tasks, particularly outperforming state-of-the-art algorithms in terms of safety.
This paper considers the problem of solving constrained reinforcement learning problems with anytime guarantees, meaning that the algorithmic solution returns a safe policy regardless of when it is terminated. Drawing inspiration from anytime constrained optimization, we introduce Reinforcement Learning-based Safe Gradient Flow (RL-SGF), an on-policy algorithm which employs estimates of the value functions and their respective gradients associated with the objective and safety constraints for the current policy, and updates the policy parameters by solving a convex quadratically constrained quadratic program. We show that if the estimates are computed with a sufficiently large number of episodes (for which we provide an explicit bound), safe policies are updated to safe policies with a probability higher than a prescribed tolerance. We also show that iterates asymptotically converge to a neighborhood of a KKT point, whose size can be arbitrarily reduced by refining the estimates of the value function and their gradients. We illustrate the performance of RL-SGF in a navigation example.
Achieving safe autonomous navigation systems is critical for deploying robots in dynamic and uncertain real-world environments. In this paper, we propose a hierarchical control framework leveraging neural network verification techniques to design control barrier functions (CBFs) and policy correction mechanisms that ensure safe reinforcement learning navigation policies. Our approach relies on probabilistic enumeration to identify unsafe regions of operation, which are then used to construct a safe CBF-based control layer applicable to arbitrary policies. We validate our framework both in simulation and on a real robot, using a standard mobile robot benchmark and a highly dynamic aquatic environmental monitoring task. These experiments demonstrate the ability of the proposed solution to correct unsafe actions while preserving efficient navigation behavior. Our results show the promise of developing hierarchical verification-based systems to enable safe and robust navigation behaviors in complex scenarios.
Safe Reinforcement Learning (Safe RL) aims to ensure safety when an RL agent conducts learning by interacting with real-world environments where improper actions can induce high costs or lead to severe consequences. In this paper, we propose a novel Safe Skill Planning (SSkP) approach to enhance effective safe RL by exploiting auxiliary offline demonstration data. SSkP involves a two-stage process. First, we employ PU learning to learn a skill risk predictor from the offline demonstration data. Then, based on the learned skill risk predictor, we develop a novel risk planning process to enhance online safe RL and learn a risk-averse safe policy efficiently through interactions with the online RL environment, while simultaneously adapting the skill risk predictor to the environment. We conduct experiments in several benchmark robotic simulation environments. The experimental results demonstrate that the proposed approach consistently outperforms previous state-of-the-art safe RL methods.
In safe reinforcement learning (RL), auxiliary safety costs are used to align the agent to safe decision making. In practice, safety constraints, including cost functions and budgets, are unknown or hard to specify, as it requires anticipation of all possible unsafe behaviors. We therefore address a general setting where the true safety definition is unknown, and has to be learned from sparsely labeled data. Our key contributions are: first, we design a safety model that performs credit assignment to estimate each decision step's impact on the overall safety using a dataset of diverse trajectories and their corresponding binary safety labels (i.e., whether the corresponding trajectory is safe/unsafe). Second, we illustrate the architecture of our safety model to demonstrate its ability to learn a separate safety score for each timestep. Third, we reformulate the safe RL problem using the proposed safety model and derive an effective algorithm to optimize a safe yet rewarding policy. Finally, our empirical results corroborate our findings and show that this approach is effective in satisfying unknown safety definition, and scalable to various continuous control tasks.
Generalizable alignment is a core challenge for deploying Large Language Models (LLMs) safely in real-world NLP applications. Current alignment methods, including Reinforcement Learning from Human Feedback (RLHF), often fail to guarantee constraint satisfaction outside their training distribution due to their reliance on implicit, post-hoc preferences. Inspired by a paradigm shift to first curate data before tuning, we introduce a new framework for safe language alignment that learns natural language constraints from positive and negative demonstrations as a primary step. From inferring both a task-specific reward function and latent constraint functions, our approach fosters adaptation to novel safety requirements and robust generalization under domain shifts and adversarial inputs. We formalize the framework within a Constrained Markov Decision Process (CMDP) and validate it via a text-based navigation environment, demonstrating safe adaptation to changing danger zones. Our experiments show fewer violations upon domain shift when following a safe navigation path, and we achieve zero violations by applying learned constraints to a distilled BERT model as a fine-tuning technique. This work offers a promising path toward building safety-critical and more generalizable LLMs for practical NLP settings.
Safe Reinforcement Learning (Safe RL) aims to train an RL agent to maximize its performance in real-world environments while adhering to safety constraints, as exceeding safety violation limits can result in severe consequences. In this paper, we propose a novel safe RL approach called Safety Modulated Policy Optimization (SMPO), which enables safe policy function learning within the standard policy optimization framework through safety modulated rewards. In particular, we consider safety violation costs as feedback from the RL environments that are parallel to the standard awards, and introduce a Q-cost function as safety critic to estimate expected future cumulative costs. Then we propose to modulate the rewards using a cost-aware weighting function, which is carefully designed to ensure the safety limits based on the estimation of the safety critic, while maximizing the expected rewards. The policy function and the safety critic are simultaneously learned through gradient descent during online interactions with the environment. We conduct experiments using multiple RL environments and the experimental results demonstrate that our method outperforms several classic and state-of-the-art comparison methods in terms of overall safe RL performance.
Ensuring safety in reinforcement learning (RL)-based robotic systems is a critical challenge, especially in contact-rich tasks within unstructured environments. While the state-of-the-art safe RL approaches mitigate risks through safe exploration or high-level recovery mechanisms, they often overlook low-level execution safety, where reflexive responses to potential hazards are crucial. Similarly, variable impedance control (VIC) enhances safety by adjusting the robot's mechanical response, yet lacks a systematic way to adapt parameters, such as stiffness and damping throughout the task. In this paper, we propose Bresa, a Bio-inspired Reflexive Hierarchical Safe RL method inspired by biological reflexes. Our method decouples task learning from safety learning, incorporating a safety critic network that evaluates action risks and operates at a higher frequency than the task solver. Unlike existing recovery-based methods, our safety critic functions at a low-level control layer, allowing real-time intervention when unsafe conditions arise. The task-solving RL policy, running at a lower frequency, focuses on high-level planning (decision-making), while the safety critic ensures instantaneous safety corrections. We validate Bresa on multiple tasks including a contact-rich robotic task, demonstrating its reflexive ability to enhance safety, and adaptability in unforeseen dynamic environments. Our results show that Bresa outperforms the baseline, providing a robust and reflexive safety mechanism that bridges the gap between high-level planning and low-level execution. Real-world experiments and supplementary material are available at project website https://jack-sherman01.github.io/Bresa.
Multimodal large language models (MLLMs) are critical for developing general-purpose AI assistants, yet they face growing safety risks. How can we ensure that MLLMs are safely aligned to prevent undesired behaviors such as discrimination, misinformation, or violations of ethical standards? In a further step, we need to explore how to fine-tune MLLMs to enhance reasoning performance while ensuring they satisfy safety constraints. Fundamentally, this can be formulated as a min-max optimization problem. In this study, we propose Safe RLHF-V, the first multimodal safety alignment framework that jointly optimizes helpfulness and safety using separate multimodal reward and cost models within a Lagrangian-based constrained optimization framework. Given that there is a lack of preference datasets that separate helpfulness and safety in multimodal scenarios, we introduce BeaverTails-V, the first open-source dataset with dual preference annotations for helpfulness and safety, along with multi-level safety labels (minor, moderate, severe). Additionally, we design a Multi-level Guardrail System to proactively defend against unsafe queries and adversarial attacks. By applying the Beaver-Guard-V moderation for 5 rounds of filtering and re-generation on the precursor model, the overall safety of the upstream model is significantly improved by an average of 40.9%. Experimental results demonstrate that fine-tuning different MLLMs with Safe RLHF can effectively enhance model helpfulness while ensuring improved safety. Specifically, Safe RLHF-V improves model safety by 34.2% and helpfulness by 34.3%. All of datasets, models, and code can be found at https://github.com/SafeRLHF-V to support the safety development of MLLMs and reduce potential societal risks.
The driving risk field is applicable to more complex driving scenarios, providing new approaches for safety decision-making and active vehicle control in intricate environments. However, existing research often overlooks the driving risk field and fails to consider the impact of risk distribution within drivable areas on trajectory planning, which poses challenges for enhancing safety. This paper proposes a trajectory planning method for intelligent vehicles based on the risk reachable set to further improve the safety of trajectory planning. First, we construct the reachable set incorporating the driving risk field to more accurately assess and avoid potential risks in drivable areas. Then, the initial trajectory is generated based on safe reinforcement learning and projected onto the reachable set. Finally, we introduce a trajectory planning method based on a constrained iterative quadratic regulator to optimize the initial solution, ensuring that the planned trajectory achieves optimal comfort, safety, and efficiency. We conduct simulation tests of trajectory planning in high-speed lane-changing scenarios. The results indicate that the proposed method can guarantee trajectory comfort and driving efficiency, with the generated trajectory situated outside high-risk boundaries, thereby ensuring vehicle safety during operation.
In real-life scenarios, a Reinforcement Learning (RL) agent aiming to maximise their reward, must often also behave in a safe manner, including at training time. Thus, much attention in recent years has been given to Safe RL, where an agent aims to learn an optimal policy among all policies that satisfy a given safety constraint. However, strict safety guarantees are often provided through approaches based on linear programming, and thus have limited scaling. In this paper we present a new, scalable method, which enjoys strict formal guarantees for Safe RL, in the case where the safety dynamics of the Markov Decision Process (MDP) are known, and safety is defined as an undiscounted probabilistic avoidance property. Our approach is based on state-augmentation of the MDP, and on the design of a shield that restricts the actions available to the agent. We show that our approach provides a strict formal safety guarantee that the agent stays safe at training and test time. Furthermore, we demonstrate that our approach is viable in practice through experimental evaluation.
Reinforcement learning-based mapless navigation holds significant potential. However, it faces challenges in indoor environments with local minima area. This paper introduces a safe mapless navigation framework utilizing hierarchical reinforcement learning (HRL) to enhance navigation through such areas. The high-level policy creates a sub-goal to direct the navigation process. Notably, we have developed a sub-goal update mechanism that considers environment congestion, efficiently avoiding the entrapment of the robot in local minimum areas. The low-level motion planning policy, trained through safe reinforcement learning, outputs real-time control instructions based on acquired sub-goal. Specifically, to enhance the robot's environmental perception, we introduce a new obstacle encoding method that evaluates the impact of obstacles on the robot's motion planning. To validate the performance of our HRL-based navigation framework, we conduct simulations in office, home, and restaurant environments. The findings demonstrate that our HRL-based navigation framework excels in both static and dynamic scenarios. Finally, we implement the HRL-based navigation framework on a TurtleBot3 robot for physical validation experiments, which exhibits its strong generalization capabilities.
In safe reinforcement learning, agent needs to balance between exploration actions and safety constraints. Following this paradigm, domain transfer approaches learn a prior Q-function from the related environments to prevent unsafe actions. However, because of the large number of false positives, some safe actions are never executed, leading to inadequate exploration in sparse-reward environments. In this work, we aim to learn an efficient state representation to balance the exploration and safety-prefer action in a sparse-reward environment. Firstly, the image input is mapped to latent representation by an auto-encoder. A further contrastive learning objective is employed to distinguish safe and unsafe states. In the learning phase, the latent distance is used to construct an additional safety check, which allows the agent to bias the exploration if it visits an unsafe state. To verify the effectiveness of our method, the experiment is carried out in three navigation-based MiniGrid environments. The result highlights that our method can explore the environment better while maintaining a good balance between safety and efficiency.
Advancing safe autonomous systems through reinforcement learning (RL) requires robust benchmarks to evaluate performance, analyze methods, and assess agent competencies. Humans primarily rely on embodied visual perception to safely navigate and interact with their surroundings, making it a valuable capability for RL agents. However, existing vision-based 3D benchmarks only consider simple navigation tasks. To address this shortcoming, we introduce \textbf{HASARD}, a suite of diverse and complex tasks to $\textbf{HA}$rness $\textbf{SA}$fe $\textbf{R}$L with $\textbf{D}$oom, requiring strategic decision-making, comprehending spatial relationships, and predicting the short-term future. HASARD features three difficulty levels and two action spaces. An empirical evaluation of popular baseline methods demonstrates the benchmark's complexity, unique challenges, and reward-cost trade-offs. Visualizing agent navigation during training with top-down heatmaps provides insight into a method's learning process. Incrementally training across difficulty levels offers an implicit learning curriculum. HASARD is the first safe RL benchmark to exclusively target egocentric vision-based learning, offering a cost-effective and insightful way to explore the potential and boundaries of current and future safe RL methods. The environments and baseline implementations are open-sourced at https://sites.google.com/view/hasard-bench/.
Vision-language-action models (VLAs) show potential as generalist robot policies. However, these models pose extreme safety challenges during real-world deployment, including the risk of harm to the environment, the robot itself, and humans. How can safety constraints be explicitly integrated into VLAs? We address this by exploring an integrated safety approach (ISA), systematically modeling safety requirements, then actively eliciting diverse unsafe behaviors, effectively constraining VLA policies via safe reinforcement learning, and rigorously assuring their safety through targeted evaluations. Leveraging the constrained Markov decision process (CMDP) paradigm, ISA optimizes VLAs from a min-max perspective against elicited safety risks. Thus, policies aligned through this comprehensive approach achieve the following key features: (I) effective safety-performance trade-offs, this exploration yields an 83.58% safety improvement compared to the current state-of-the-art method, while also maintaining task performance (+3.85%). (II) strong safety assurance, with the ability to mitigate long-tail risks and handle extreme failure scenarios. (III) robust generalization of learned safety behaviors to various out-of-distribution perturbations. Our data, models and newly proposed benchmark environment are available at https://pku-safevla.github.io.