Designing controllers that accomplish tasks while guaranteeing safety constraints remains a significant challenge. We often want an agent to perform well in a nominal task, such as environment exploration, while ensuring it can avoid unsafe states and return to a desired target by a specific time. In particular we are motivated by the setting of safe, efficient, hands-off training for reinforcement learning in the real world. By enabling a robot to safely and autonomously reset to a desired region (e.g., charging stations) without human intervention, we can enhance efficiency and facilitate training. Safety filters, such as those based on control barrier functions, decouple safety from nominal control objectives and rigorously guarantee safety. Despite their success, constructing these functions for general nonlinear systems with control constraints and system uncertainties remains an open problem. This paper introduces a safety filter obtained from the value function associated with the reach-avoid problem. The proposed safety filter minimally modifies the nominal controller while avoiding unsafe regions and guiding the system back to the desired target set. By preserving policy performance while allowing safe resetting, we enable efficient hands-off reinforcement learning and advance the feasibility of safe training for real world robots. We demonstrate our approach using a modified version of soft actor-critic to safely train a swing-up task on a modified cartpole stabilization problem.

Traffic signal control (TSC) is an important and widely studied direction. Recently, reinforcement learning (RL) methods have been used to solve TSC problems and achieve superior performance over conventional TSC methods. However, applying RL methods to the real world is challenging due to the huge cost of experiments in real-world traffic environments. One possible solution is TSC domain adaptation, which adapts trained models to target environments and reduces the number of interactions and the training cost. However, existing TSC domain adaptation methods still face two major issues: the lack of consideration for differences across cities and the low utilization of multi-city data. To solve aforementioned issues, we propose an approach named Adaptive Modularized Model (AMM). By modularizing TSC problems and network models, we overcome the challenge of possible changes in environmental observations. We also aggregate multi-city experience through meta-learning. We conduct extensive experiments on different cities and show that AMM can achieve excellent performance with limited interactions in target environments and outperform existing methods. We also demonstrate the feasibility and generalizability of our method.

Humans rely on high-level meta-representations to engage in abstract reasoning. In complex cognitive tasks, these meta-representations help individuals abstract general rules from experience. However, constructing such meta-representations from high-dimensional observations remains a longstanding challenge for reinforcement learning agents. For instance, a well-trained agent often fails to generalize to even minor variations of the same task, such as changes in background color, while humans can easily handle. In this paper, we build a bridge between meta-representation and generalization, showing that generalization performance benefits from meta-representation learning. We also hypothesize that deep mutual learning (DML) among agents can help them converge to meta-representations. Empirical results provide support for our theory and hypothesis. Overall, this work provides a new perspective on the generalization of deep reinforcement learning.

Reinforcement learning (RL) enables an agent interacting with an unknown MDP $M$ to optimise its behaviour by observing transitions sampled from $M$. A natural entity that emerges in the agent's reasoning is $\widehat{M}$, the maximum likelihood estimate of $M$ based on the observed transitions. The well-known \textit{certainty-equivalence} method (CEM) dictates that the agent update its behaviour to $\widehat{\pi}$, which is an optimal policy for $\widehat{M}$. Not only is CEM intuitive, it has been shown to enjoy minimax-optimal sample complexity in some regions of the parameter space for PAC RL with a generative model~\citep{Agarwal2020GenModel}. A seemingly unrelated algorithm is the ``trajectory tree method'' (TTM)~\citep{Kearns+MN:1999}, originally developed for efficient decision-time planning in large POMDPs. This paper presents a theoretical investigation that stems from the surprising finding that CEM may indeed be viewed as an application of TTM. The qualitative benefits of this view are (1) new and simple proofs of sample complexity upper bounds for CEM, in fact under a (2) weaker assumption on the rewards than is prevalent in the current literature. Our analysis applies to both non-stationary and stationary MDPs. Quantitatively, we obtain (3) improvements in the sample-complexity upper bounds for CEM both for non-stationary and stationary MDPs, in the regime that the ``mistake probability'' $\delta$ is small. Additionally, we show (4) a lower bound on the sample complexity for finite-horizon MDPs, which establishes the minimax-optimality of our upper bound for non-stationary MDPs in the small-$\delta$ regime.

Distributionally robust offline reinforcement learning (RL) aims to find a policy that performs the best under the worst environment within an uncertainty set using an offline dataset collected from a nominal model. While recent advances in robust RL focus on Markov decision processes (MDPs), robust non-Markovian RL is limited to planning problem where the transitions in the uncertainty set are known. In this paper, we study the learning problem of robust offline non-Markovian RL. Specifically, when the nominal model admits a low-rank structure, we propose a new algorithm, featuring a novel dataset distillation and a lower confidence bound (LCB) design for robust values under different types of the uncertainty set. We also derive new dual forms for these robust values in non-Markovian RL, making our algorithm more amenable to practical implementation. By further introducing a novel type-I concentrability coefficient tailored for offline low-rank non-Markovian decision processes, we prove that our algorithm can find an $\epsilon$-optimal robust policy using $O(1/\epsilon^2)$ offline samples. Moreover, we extend our algorithm to the case when the nominal model does not have specific structure. With a new type-II concentrability coefficient, the extended algorithm also enjoys polynomial sample efficiency under all different types of the uncertainty set.

In this paper, we propose a novel method for learning reward functions directly from offline demonstrations. Unlike traditional inverse reinforcement learning (IRL), our approach decouples the reward function from the learner's policy, eliminating the adversarial interaction typically required between the two. This results in a more stable and efficient training process. Our reward function, called \textit{SR-Reward}, leverages successor representation (SR) to encode a state based on expected future states' visitation under the demonstration policy and transition dynamics. By utilizing the Bellman equation, SR-Reward can be learned concurrently with most reinforcement learning (RL) algorithms without altering the existing training pipeline. We also introduce a negative sampling strategy to mitigate overestimation errors by reducing rewards for out-of-distribution data, thereby enhancing robustness. This strategy inherently introduces a conservative bias into RL algorithms that employ the learned reward. We evaluate our method on the D4RL benchmark, achieving competitive results compared to offline RL algorithms with access to true rewards and imitation learning (IL) techniques like behavioral cloning. Moreover, our ablation studies on data size and quality reveal the advantages and limitations of SR-Reward as a proxy for true rewards.

This study is motivated by optimal control applications that exhibit high complexity, high uncertainty, and very limited data [Wang et al., 2024, Zheng et al., 2023, Plotkin et al., 2017, Mirasol, 2017]. In particular, all of these challenges are present in the domain of biopharmaceutical manufacturing, used for production of essential life-saving treatments for severe and chronic diseases, including cancers, autoimmune disorders, metabolic diseases, genetic disorders, and infectious diseases such as COVID-19 [Zahavi and Weiner, 2020, Teo, 2022]. Using cells as factories, biomanufacturing involves hundreds of biological, physical, and chemical factors dynamically interacting with each other at molecular, cellular, and macroscopic levels and impacting production outcomes. Due to the complexity of these mechanisms, it is quite difficult to control production safely and effectively, especially in the presence of very limited data. Digital twins have proven very useful in guiding the control of complex physical systems [Tao et al., 2018].

In domains such as finance, healthcare, and robotics, managing worst-case scenarios is critical, as failure to do so can lead to catastrophic outcomes. Distributional Reinforcement Learning (DRL) provides a natural framework to incorporate risk sensitivity into decision-making processes. However, existing approaches face two key limitations: (1) the use of fixed risk measures at each decision step often results in overly conservative policies, and (2) the interpretation and theoretical properties of the learned policies remain unclear. While optimizing a static risk measure addresses these issues, its use in the DRL framework has been limited to the simple static CVaR risk measure. In this paper, we present a novel DRL algorithm with convergence guarantees that optimizes for a broader class of static Spectral Risk Measures (SRM). Additionally, we provide a clear interpretation of the learned policy by leveraging the distribution of returns in DRL and the decomposition of static coherent risk measures. Extensive experiments demonstrate that our model learns policies aligned with the SRM objective, and outperforms existing risk-neutral and risk-sensitive DRL models in various settings.

Humanoid robots have great potential to perform various human-level skills. These skills involve locomotion, manipulation, and cognitive capabilities. Driven by advances in machine learning and the strength of existing model-based approaches, these capabilities have progressed rapidly, but often separately. Therefore, a timely overview of current progress and future trends in this fast-evolving field is essential. This survey first summarizes the model-based planning and control that have been the backbone of humanoid robotics for the past three decades. We then explore emerging learning-based methods, with a focus on reinforcement learning and imitation learning that enhance the versatility of loco-manipulation skills. We examine the potential of integrating foundation models with humanoid embodiments, assessing the prospects for developing generalist humanoid agents. In addition, this survey covers emerging research for whole-body tactile sensing that unlocks new humanoid skills that involve physical interactions. The survey concludes with a discussion of the challenges and future trends.

AlphaGo's astonishing performance has ignited an explosive interest in developing deep reinforcement learning (DRL) for numerous real-world applications, such as intelligent robotics. However, the often prohibitive complexity of DRL stands at the odds with the required real-time control and constrained resources in many DRL applications, limiting the great potential of DRL powered intelligent devices. While substantial efforts have been devoted to compressing other deep learning models, existing works barely touch the surface of compressing DRL. In this work, we first identify that there exists an optimal model size of DRL that can maximize both the test scores and efficiency, motivating the need for task-specific DRL agents. We therefore propose an Auto-Agent-Distiller (A2D) framework, which to our best knowledge is the first neural architecture search (NAS) applied to DRL to automatically search for the optimal DRL agents for various tasks that optimize both the test scores and efficiency. Specifically, we demonstrate that vanilla NAS can easily fail in searching for the optimal agents, due to its resulting high variance in DRL training stability, and then develop a novel distillation mechanism to distill the knowledge from both the teacher agent's actor and critic to stabilize the searching process and improve the searched agents' optimality. Extensive experiments and ablation studies consistently validate our findings and the advantages and general applicability of our A2D, outperforming manually designed DRL in both the test scores and efficiency. All the codes will be released upon acceptance.