Reinforcement learning (RL) has achieved tremendous success in many complex decision making tasks. When it comes to deploying RL in the real world, safety concerns are usually raised, leading to a growing demand for safe RL algorithms, such as in autonomous driving and robotics scenarios. While safety control has a long history, the study of safe RL algorithms is still in the early stages. To establish a good foundation for future research in this thread, in this paper, we provide a review for safe RL from the perspectives of methods, theory and applications. Firstly, we review the progress of safe RL from five dimensions and come up with five problems that are crucial for safe RL being deployed in real-world applications, coined as "2H3W". Secondly, we analyze the theory and algorithm progress from the perspectives of answering the "2H3W" problems. Then, the sample complexity of safe RL methods is reviewed and discussed, followed by an introduction of the applications and benchmarks of safe RL algorithms. Finally, we open the discussion of the challenging problems in safe RL, hoping to inspire more future research on this thread. To advance the study of safe RL algorithms, we release a benchmark suite, an open-sourced repository containing the implementations of major safe RL algorithms, along with tutorials at the link: https://github.com/chauncygu/Safe-Reinforcement-Learning-Baselines.git.

Recently, several studies (Zhou et al., 2021a; Zhang et al., 2021b; Kim et al., 2021; Zhou and Gu, 2022) have provided variance-dependent regret bounds for linear contextual bandits, which interpolates the regret for the worst-case regime and the deterministic reward regime. However, these algorithms are either computationally intractable or unable to handle unknown variance of the noise. In this paper, we present a novel solution to this open problem by proposing the first computationally efficient algorithm for linear bandits with heteroscedastic noise. Our algorithm is adaptive to the unknown variance of noise and achieves an $\tilde{O}(d \sqrt{\sum_{k = 1}^K \sigma_k^2} + d)$ regret, where $\sigma_k^2$ is the variance of the noise at the round $k$, $d$ is the dimension of the contexts and $K$ is the total number of rounds. Our results are based on an adaptive variance-aware confidence set enabled by a new Freedman-type concentration inequality for self-normalized martingales and a multi-layer structure to stratify the context vectors into different layers with different uniform upper bounds on the uncertainty. Furthermore, our approach can be extended to linear mixture Markov decision processes (MDPs) in reinforcement learning. We propose a variance-adaptive algorithm for linear mixture MDPs, which achieves a problem-dependent horizon-free regret bound that can gracefully reduce to a nearly constant regret for deterministic MDPs. Unlike existing nearly minimax optimal algorithms for linear mixture MDPs, our algorithm does not require explicit variance estimation of the transitional probabilities or the use of high-order moment estimators to attain horizon-free regret. We believe the techniques developed in this paper can have independent value for general online decision making problems.

Model-based reinforcement learning uses models to plan, where the predictions and policies of an agent can be improved by using more computation without additional data from the environment, thereby improving sample efficiency. However, learning accurate estimates of the model is hard. Subsequently, the natural question is whether we can get similar benefits as planning with model-free methods. Experience replay is an essential component of many model-free algorithms enabling sample-efficient learning and stability by providing a mechanism to store past experiences for further reuse in the gradient computational process. Prior works have established connections between models and experience replay by planning with the latter. This involves increasing the number of times a mini-batch is sampled and used for updates at each step (amount of replay per step). We attempt to exploit this connection by doing a systematic study on the effect of varying amounts of replay per step in a well-known model-free algorithm: Deep Q-Network (DQN) in the Mountain Car environment. We empirically show that increasing replay improves DQN's sample efficiency, reduces the variation in its performance, and makes it more robust to change in hyperparameters. Altogether, this takes a step toward a better algorithm for deployment.

Task automation of surgical robot has the potentials to improve surgical efficiency. Recent reinforcement learning (RL) based approaches provide scalable solutions to surgical automation, but typically require extensive data collection to solve a task if no prior knowledge is given. This issue is known as the exploration challenge, which can be alleviated by providing expert demonstrations to an RL agent. Yet, how to make effective use of demonstration data to improve exploration efficiency still remains an open challenge. In this work, we introduce Demonstration-guided EXploration (DEX), an efficient reinforcement learning algorithm that aims to overcome the exploration problem with expert demonstrations for surgical automation. To effectively exploit demonstrations, our method estimates expert-like behaviors with higher values to facilitate productive interactions, and adopts non-parametric regression to enable such guidance at states unobserved in demonstration data. Extensive experiments on $10$ surgical manipulation tasks from SurRoL, a comprehensive surgical simulation platform, demonstrate significant improvements in the exploration efficiency and task success rates of our method. Moreover, we also deploy the learned policies to the da Vinci Research Kit (dVRK) platform to show the effectiveness on the real robot. Code is available at https://github.com/med-air/DEX.

We propose Meta-World Conditional Neural Processes (MW-CNP), a conditional world model generator that leverages sample efficiency and scalability of Conditional Neural Processes to enable an agent to sample from its own "hallucination". We intend to reduce the agent's interaction with the target environment at test time as much as possible. To reduce the number of samples required at test time, we first obtain a latent representation of the transition dynamics from a single rollout from the test environment with hidden parameters. Then, we obtain rollouts for few-shot learning by interacting with the "hallucination" generated by the meta-world model. Using the world model representation from MW-CNP, the meta-RL agent can adapt to an unseen target environment with significantly fewer samples collected from the target environment compared to the baselines. We emphasize that the agent does not have access to the task parameters throughout training and testing, and MW-CNP is trained on offline interaction data logged during meta-training.

This work aims to tackle a major challenge in offline Inverse Reinforcement Learning (IRL), namely the reward extrapolation error, where the learned reward function may fail to explain the task correctly and misguide the agent in unseen environments due to the intrinsic covariate shift. Leveraging both expert data and lower-quality diverse data, we devise a principled algorithm (namely CLARE) that solves offline IRL efficiently via integrating "conservatism" into a learned reward function and utilizing an estimated dynamics model. Our theoretical analysis provides an upper bound on the return gap between the learned policy and the expert policy, based on which we characterize the impact of covariate shift by examining subtle two-tier tradeoffs between the "exploitation" (on both expert and diverse data) and "exploration" (on the estimated dynamics model). We show that CLARE can provably alleviate the reward extrapolation error by striking the right "exploitation-exploration" balance therein. Extensive experiments corroborate the significant performance gains of CLARE over existing state-of-the-art algorithms on MuJoCo continuous control tasks (especially with a small offline dataset), and the learned reward is highly instructive for further learning (source code). The primary objective of Inverse Reinforcement Learning (IRL) is to learn a reward function from demonstrations (Arora & Doshi, 2021; Russell, 1998). In general, conventional IRL methods rely on extensive online trials and errors that can be costly or require a fully known transition model (Abbeel & Ng, 2004; Ratliff et al., 2006; Ziebart et al., 2008; Syed & Schapire, 2007; Boularias et al., 2011; Osa et al., 2018), struggling to scale in many real-world applications. To tackle this problem, this paper studies offline IRL, with focus on learning from a previously collected dataset without online interaction with the environment.

We study the problem of generating control laws for systems with unknown dynamics. Our approach is to represent the controller and the value function with neural networks, and to train them using loss functions adapted from the Hamilton-Jacobi-Bellman (HJB) equations. In the absence of a known dynamics model, our method first learns the state transitions from data collected by interacting with the system in an offline process. The learned transition function is then integrated to the HJB equations and used to forward simulate the control signals produced by our controller in a feedback loop. In contrast to trajectory optimization methods that optimize the controller for a single initial state, our controller can generate near-optimal control signals for initial states from a large portion of the state space. Compared to recent model-based reinforcement learning algorithms, we show that our method is more sample efficient and trains faster by an order of magnitude. We demonstrate our method in a number of tasks, including the control of a quadrotor with 12 state variables.

Deep Policy Gradient (PG) algorithms employ value networks to drive the learning of parameterized policies and reduce the variance of the gradient estimates. However, value function approximation gets stuck in local optima and struggles to fit the actual return, limiting the variance reduction efficacy and leading policies to sub-optimal performance. This paper focuses on improving value approximation and analyzing the effects on Deep PG primitives such as value prediction, variance reduction, and correlation of gradient estimates with the true gradient. To this end, we introduce a Value Function Search that employs a population of perturbed value networks to search for a better approximation. Our framework does not require additional environment interactions, gradient computations, or ensembles, providing a computationally inexpensive approach to enhance the supervised learning task on which value networks train. Crucially, we show that improving Deep PG primitives results in improved sample efficiency and policies with higher returns using common continuous control benchmark domains.

Cost functions are commonly employed in Safe Deep Reinforcement Learning (DRL). However, the cost is typically encoded as an indicator function due to the difficulty of quantifying the risk of policy decisions in the state space. Such an encoding requires the agent to visit numerous unsafe states to learn a cost-value function to drive the learning process toward safety. Hence, increasing the number of unsafe interactions and decreasing sample efficiency. In this paper, we investigate an alternative approach that uses domain knowledge to quantify the risk in the proximity of such states by defining a violation metric. This metric is computed by verifying task-level properties, shaped as input-output conditions, and it is used as a penalty to bias the policy away from unsafe states without learning an additional value function. We investigate the benefits of using the violation metric in standard Safe DRL benchmarks and robotic mapless navigation tasks. The navigation experiments bridge the gap between Safe DRL and robotics, introducing a framework that allows rapid testing on real robots. Our experiments show that policies trained with the violation penalty achieve higher performance over Safe DRL baselines and significantly reduce the number of visited unsafe states.

While reinforcement learning (RL) shows promising results in automated learning of behavioral skills, it is still not enough to solve a challenging uninformed search problem where the desired behavior and rewards are sparsely observed. Some techniques tackle this problem by utilizing the shaped reward (Hartikainen et al., 2019) or combining representation learning for efficient exploration (Ghosh et al., 2018). But, these not only become prohibitively time-consuming in terms of the required human efforts, but also require significant domain knowledge for shaping the reward or designing the task-specific representation learning objective. What if we could design the algorithm that automatically progresses toward the desired behavior without any domain knowledge and human efforts, while distilling the experiences into the general purpose policy? An effective scheme for designing such an algorithm is one that learns on a tailored sequence of curriculum goals, allowing the agent to autonomously practice the intermediate tasks. However, a fundamental challenge is that proposing the curriculum goal to the agent is intimately connected to the efficient desired outcome-directed exploration and vice versa. If the curriculum generation is ineffective for recognizing frontier parts of the explored and feasible areas, an efficient exploration toward the desired outcome states cannot be performed. Even though some prior works propose to modify the curriculum distribution into a uniform one over the feasible state space (Pong et al., 2019; Klink et al., 2022) or generate a curriculum based on the level of difficulty (Florensa et al., 2018; Sukhbaatar et al., 2017), most of these methods show slow curriculum progress due to the process of skewing the curriculum distribution toward the uniform one rather than the frontier of the explored region or the properties that are susceptible to focusing on infeasible goals where the agent's capability stagnates in the intermediate level of difficulty.