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.
Safe navigation is essential for autonomous systems operating in hazardous environments. Traditional planning methods excel at long-horizon tasks but rely on a predefined graph with fixed distance metrics. In contrast, safe Reinforcement Learning (RL) can learn complex behaviors without relying on manual heuristics but fails to solve long-horizon tasks, particularly in goal-conditioned and multi-agent scenarios. In this paper, we introduce a novel method that integrates the strengths of both planning and safe RL. Our method leverages goal-conditioned RL and safe RL to learn a goal-conditioned policy for navigation while concurrently estimating cumulative distance and safety levels using learned value functions via an automated self-training algorithm. By constructing a graph with states from the replay buffer, our method prunes unsafe edges and generates a waypoint-based plan that the agent follows until reaching its goal, effectively balancing faster and safer routes over extended distances. Utilizing this unified high-level graph and a shared low-level goal-conditioned safe RL policy, we extend this approach to address the multi-agent safe navigation problem. In particular, we leverage Conflict-Based Search (CBS) to create waypoint-based plans for multiple agents allowing for their safe navigation over extended horizons. This integration enhances the scalability of goal-conditioned safe RL in multi-agent scenarios, enabling efficient coordination among agents. Extensive benchmarking against state-of-the-art baselines demonstrates the effectiveness of our method in achieving distance goals safely for multiple agents in complex and hazardous environments. Our code and further details about or work is available at https://safe-visual-mapf-mers.csail.mit.edu/.
The primary goal of reinforcement learning is to develop decision-making policies that prioritize optimal performance, frequently without considering risk or safety. In contrast, safe reinforcement learning seeks to reduce or avoid unsafe states. This letter introduces a risk-averse temporal difference algorithm that uses optimal transport theory to direct the agent toward predictable behavior. By incorporating a risk indicator, the agent learns to favor actions with predictable consequences. We evaluate the proposed algorithm in several case studies and show its effectiveness in the presence of uncertainty. The results demonstrate that our method reduces the frequency of visits to risky states while preserving performance. A Python implementation of the algorithm is available at https:// github.com/SAILRIT/Risk-averse-TD-Learning.
Offline safe reinforcement learning (OSRL) involves learning a decision-making policy to maximize rewards from a fixed batch of training data to satisfy pre-defined safety constraints. However, adapting to varying safety constraints during deployment without retraining remains an under-explored challenge. To address this challenge, we introduce constraint-adaptive policy switching (CAPS), a wrapper framework around existing offline RL algorithms. During training, CAPS uses offline data to learn multiple policies with a shared representation that optimize different reward and cost trade-offs. During testing, CAPS switches between those policies by selecting at each state the policy that maximizes future rewards among those that satisfy the current cost constraint. Our experiments on 38 tasks from the DSRL benchmark demonstrate that CAPS consistently outperforms existing methods, establishing a strong wrapper-based baseline for OSRL. The code is publicly available at https://github.com/yassineCh/CAPS.
Offline safe reinforcement learning (RL) has emerged as a promising approach for learning safe behaviors without engaging in risky online interactions with the environment. Most existing methods in offline safe RL rely on cost constraints at each time step (derived from global cost constraints) and this can result in either overly conservative policies or violation of safety constraints. In this paper, we propose to learn a policy that generates desirable trajectories and avoids undesirable trajectories. To be specific, we first partition the pre-collected dataset of state-action trajectories into desirable and undesirable subsets. Intuitively, the desirable set contains high reward and safe trajectories, and undesirable set contains unsafe trajectories and low-reward safe trajectories. Second, we learn a policy that generates desirable trajectories and avoids undesirable trajectories, where (un)desirability scores are provided by a classifier learnt from the dataset of desirable and undesirable trajectories. This approach bypasses the computational complexity and stability issues of a min-max objective that is employed in existing methods. Theoretically, we also show our approach's strong connections to existing learning paradigms involving human feedback. Finally, we extensively evaluate our method using the DSRL benchmark for offline safe RL. Empirically, our method outperforms competitive baselines, achieving higher rewards and better constraint satisfaction across a wide variety of benchmark tasks.
Safe reinforcement learning (RL) is a popular and versatile paradigm to learn reward-maximizing policies with safety guarantees. Previous works tend to express the safety constraints in an expectation form due to the ease of implementation, but this turns out to be ineffective in maintaining safety constraints with high probability. To this end, we move to the quantile-constrained RL that enables a higher level of safety without any expectation-form approximations. We directly estimate the quantile gradients through sampling and provide the theoretical proofs of convergence. Then a tilted update strategy for quantile gradients is implemented to compensate the asymmetric distributional density, with a direct benefit of return performance. Experiments demonstrate that the proposed model fully meets safety requirements (quantile constraints) while outperforming the state-of-the-art benchmarks with higher return.
In safe offline reinforcement learning (RL), the objective is to develop a policy that maximizes cumulative rewards while strictly adhering to safety constraints, utilizing only offline data. Traditional methods often face difficulties in balancing these constraints, leading to either diminished performance or increased safety risks. We address these issues with a novel approach that begins by learning a conservatively safe policy through the use of Conditional Variational Autoencoders, which model the latent safety constraints. Subsequently, we frame this as a Constrained Reward-Return Maximization problem, wherein the policy aims to optimize rewards while complying with the inferred latent safety constraints. This is achieved by training an encoder with a reward-Advantage Weighted Regression objective within the latent constraint space. Our methodology is supported by theoretical analysis, including bounds on policy performance and sample complexity. Extensive empirical evaluation on benchmark datasets, including challenging autonomous driving scenarios, demonstrates that our approach not only maintains safety compliance but also excels in cumulative reward optimization, surpassing existing methods. Additional visualizations provide further insights into the effectiveness and underlying mechanisms of our approach.
The high costs and risks involved in extensive environment interactions hinder the practical application of current online safe reinforcement learning (RL) methods. While offline safe RL addresses this by learning policies from static datasets, the performance therein is usually limited due to reliance on data quality and challenges with out-of-distribution (OOD) actions. Inspired by recent successes in offline-to-online (O2O) RL, it is crucial to explore whether offline safe RL can be leveraged to facilitate faster and safer online policy learning, a direction that has yet to be fully investigated. To fill this gap, we first demonstrate that naively applying existing O2O algorithms from standard RL would not work well in the safe RL setting due to two unique challenges: \emph{erroneous Q-estimations}, resulted from offline-online objective mismatch and offline cost sparsity, and \emph{Lagrangian mismatch}, resulted from difficulties in aligning Lagrange multipliers between offline and online policies. To address these challenges, we introduce \textbf{Marvel}, a novel framework for O2O safe RL, comprising two key components that work in concert: \emph{Value Pre-Alignment} to align the Q-functions with the underlying truth before online learning, and \emph{Adaptive PID Control} to effectively adjust the Lagrange multipliers during online finetuning. Extensive experiments demonstrate that Marvel significantly outperforms existing baselines in both reward maximization and safety constraint satisfaction. By introducing the first policy-finetuning based framework for O2O safe RL, which is compatible with many offline and online safe RL methods, our work has the great potential to advance the field towards more efficient and practical safe RL solutions.
Safe reinforcement learning (RL) is crucial for real-world applications, and multi-agent interactions introduce additional safety challenges. While Probabilistic Logic Shields (PLS) has been a powerful proposal to enforce safety in single-agent RL, their generalizability to multi-agent settings remains unexplored. In this paper, we address this gap by conducting extensive analyses of PLS within decentralized, multi-agent environments, and in doing so, propose Shielded Multi-Agent Reinforcement Learning (SMARL) as a general framework for steering MARL towards norm-compliant outcomes. Our key contributions are: (1) a novel Probabilistic Logic Temporal Difference (PLTD) update for shielded, independent Q-learning, which incorporates probabilistic constraints directly into the value update process; (2) a probabilistic logic policy gradient method for shielded PPO with formal safety guarantees for MARL; and (3) comprehensive evaluation across symmetric and asymmetrically shielded $n$-player game-theoretic benchmarks, demonstrating fewer constraint violations and significantly better cooperation under normative constraints. These results position SMARL as an effective mechanism for equilibrium selection, paving the way toward safer, socially aligned multi-agent systems.
Safety is one of the key issues preventing the deployment of reinforcement learning techniques in real-world robots. While most approaches in the Safe Reinforcement Learning area do not require prior knowledge of constraints and robot kinematics and rely solely on data, it is often difficult to deploy them in complex real-world settings. Instead, model-based approaches that incorporate prior knowledge of the constraints and dynamics into the learning framework have proven capable of deploying the learning algorithm directly on the real robot. Unfortunately, while an approximated model of the robot dynamics is often available, the safety constraints are task-specific and hard to obtain: they may be too complicated to encode analytically, too expensive to compute, or it may be difficult to envision a priori the long-term safety requirements. In this paper, we bridge this gap by extending the safe exploration method, ATACOM, with learnable constraints, with a particular focus on ensuring long-term safety and handling of uncertainty. Our approach is competitive or superior to state-of-the-art methods in final performance while maintaining safer behavior during training.
Recent advancements in machine learning based energy management approaches, specifically reinforcement learning with a safety layer (OptLayerPolicy) and a metaheuristic algorithm generating a decision tree control policy (TreeC), have shown promise. However, their effectiveness has only been demonstrated in computer simulations. This paper presents the real-world validation of these methods, comparing against model predictive control and simple rule-based control benchmark. The experiments were conducted on the electrical installation of 4 reproductions of residential houses, which all have their own battery, photovoltaic and dynamic load system emulating a non-controllable electrical load and a controllable electric vehicle charger. The results show that the simple rules, TreeC, and model predictive control-based methods achieved similar costs, with a difference of only 0.6%. The reinforcement learning based method, still in its training phase, obtained a cost 25.5\% higher to the other methods. Additional simulations show that the costs can be further reduced by using a more representative training dataset for TreeC and addressing errors in the model predictive control implementation caused by its reliance on accurate data from various sources. The OptLayerPolicy safety layer allows safe online training of a reinforcement learning agent in the real-world, given an accurate constraint function formulation. The proposed safety layer method remains error-prone, nonetheless, it is found beneficial for all investigated methods. The TreeC method, which does require building a realistic simulation for training, exhibits the safest operational performance, exceeding the grid limit by only 27.1 Wh compared to 593.9 Wh for reinforcement learning.
Offline safe reinforcement learning (RL) aims to train a policy that satisfies constraints using a pre-collected dataset. Most current methods struggle with the mismatch between imperfect demonstrations and the desired safe and rewarding performance. In this paper, we introduce OASIS (cOnditionAl diStributIon Shaping), a new paradigm in offline safe RL designed to overcome these critical limitations. OASIS utilizes a conditional diffusion model to synthesize offline datasets, thus shaping the data distribution toward a beneficial target domain. Our approach makes compliance with safety constraints through effective data utilization and regularization techniques to benefit offline safe RL training. Comprehensive evaluations on public benchmarks and varying datasets showcase OASIS's superiority in benefiting offline safe RL agents to achieve high-reward behavior while satisfying the safety constraints, outperforming established baselines. Furthermore, OASIS exhibits high data efficiency and robustness, making it suitable for real-world applications, particularly in tasks where safety is imperative and high-quality demonstrations are scarce.
The growing complexity of power system management has led to an increased interest in reinforcement learning (RL). To validate their effectiveness, RL algorithms have to be evaluated across multiple case studies. Case study design is an arduous task requiring the consideration of many aspects, among them the influence of available forecasts and the level of decentralization in the control structure. Furthermore, vanilla RL controllers cannot themselves ensure the satisfaction of system constraints, which makes devising a safeguarding mechanism a necessary task for every case study before deploying the system. To address these shortcomings, we introduce the Python tool CommonPower, the first general framework for the modeling and simulation of power system management tailored towards machine learning. Its modular architecture enables users to focus on specific elements without having to implement a simulation environment. Another unique contribution of CommonPower is the automatic synthesis of model predictive controllers and safeguards. Beyond offering a unified interface for single-agent RL, multi-agent RL, and optimal control, CommonPower includes a training pipeline for machine-learning-based forecasters as well as a flexible mechanism for incorporating feedback of safeguards into the learning updates of RL controllers.
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%.
Empowering safe exploration of reinforcement learning (RL) agents during training is a critical challenge towards their deployment in many real-world scenarios. When prior knowledge of the domain or task is unavailable, training RL agents in unknown, \textit{black-box} environments presents an even greater safety risk. We introduce \mbox{ADVICE} (Adaptive Shielding with a Contrastive Autoencoder), a novel post-shielding technique that distinguishes safe and unsafe features of state-action pairs during training, and uses this knowledge to protect the RL agent from executing actions that yield likely hazardous outcomes. Our comprehensive experimental evaluation against state-of-the-art safe RL exploration techniques shows that ADVICE significantly reduces safety violations ($\approx\!\!50\%$) during training, with a competitive outcome reward compared to other techniques.
Learning reliably safe autonomous control is one of the core problems in trustworthy autonomy. However, training a controller that can be formally verified to be safe remains a major challenge. We introduce a novel approach for learning verified safe control policies in nonlinear neural dynamical systems while maximizing overall performance. Our approach aims to achieve safety in the sense of finite-horizon reachability proofs, and is comprised of three key parts. The first is a novel curriculum learning scheme that iteratively increases the verified safe horizon. The second leverages the iterative nature of gradient-based learning to leverage incremental verification, reusing information from prior verification runs. Finally, we learn multiple verified initial-state-dependent controllers, an idea that is especially valuable for more complex domains where learning a single universal verified safe controller is extremely challenging. Our experiments on five safe control problems demonstrate that our trained controllers can achieve verified safety over horizons that are as much as an order of magnitude longer than state-of-the-art baselines, while maintaining high reward, as well as a perfect safety record over entire episodes. Our code is available at https://github.com/jlwu002/VSRL.
In the field of safe reinforcement learning (RL), finding a balance between satisfying safety constraints and optimizing reward performance presents a significant challenge. A key obstacle in this endeavor is the estimation of safety constraints, which is typically more difficult than estimating a reward metric due to the sparse nature of the constraint signals. To address this issue, we introduce a novel framework named Feasibility Consistent Safe Reinforcement Learning (FCSRL). This framework combines representation learning with feasibility-oriented objectives to identify and extract safety-related information from the raw state for safe RL. Leveraging self-supervised learning techniques and a more learnable safety metric, our approach enhances the policy learning and constraint estimation. Empirical evaluations across a range of vector-state and image-based tasks demonstrate that our method is capable of learning a better safety-aware embedding and achieving superior performance than previous representation learning baselines.
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 RL benchmark, Omnisafe. Experimental results demonstrate that our algorithms outperform several state-of-the-art baselines in terms of balancing reward and safety optimization.
Recent progress in large language models (LLMs) has led to their widespread adoption in various domains. However, these advancements have also introduced additional safety risks and raised concerns regarding their detrimental impact on already marginalized populations. Despite growing mitigation efforts to develop safety safeguards, such as supervised safety-oriented fine-tuning and leveraging safe reinforcement learning from human feedback, multiple concerns regarding the safety and ingrained biases in these models remain. Furthermore, previous work has demonstrated that models optimized for safety often display exaggerated safety behaviors, such as a tendency to refrain from responding to certain requests as a precautionary measure. As such, a clear trade-off between the helpfulness and safety of these models has been documented in the literature. In this paper, we further investigate the effectiveness of safety measures by evaluating models on already mitigated biases. Using the case of Llama 2 as an example, we illustrate how LLMs' safety responses can still encode harmful assumptions. To do so, we create a set of non-toxic prompts, which we then use to evaluate Llama models. Through our new taxonomy of LLMs responses to users, we observe that the safety/helpfulness trade-offs are more pronounced for certain demographic groups which can lead to quality-of-service harms for marginalized populations.
Safe reinforcement learning (RL) offers advanced solutions to constrained optimal control problems. Existing studies in safe RL implicitly assume continuity in policy functions, where policies map states to actions in a smooth, uninterrupted manner; however, our research finds that in some scenarios, the feasible policy should be discontinuous or multi-valued, interpolating between discontinuous local optima can inevitably lead to constraint violations. We are the first to identify the generating mechanism of such a phenomenon, and employ topological analysis to rigorously prove the existence of policy bifurcation in safe RL, which corresponds to the contractibility of the reachable tuple. Our theorem reveals that in scenarios where the obstacle-free state space is non-simply connected, a feasible policy is required to be bifurcated, meaning its output action needs to change abruptly in response to the varying state. To train such a bifurcated policy, we propose a safe RL algorithm called multimodal policy optimization (MUPO), which utilizes a Gaussian mixture distribution as the policy output. The bifurcated behavior can be achieved by selecting the Gaussian component with the highest mixing coefficient. Besides, MUPO also integrates spectral normalization and forward KL divergence to enhance the policy's capability of exploring different modes. Experiments with vehicle control tasks show that our algorithm successfully learns the bifurcated policy and ensures satisfying safety, while a continuous policy suffers from inevitable constraint violations.