A BSTRACT In sparse reward scenarios of reinforcement learning (RL), the memory mechanism provides promising shortcuts to policy optimization by reflecting on past experiences like humans. However, current memory-based RL methods simply store and reuse high-value policies, lacking a deeper refining and filtering of diverse past experiences and hence limiting the capability of memory. In this paper, we propose AdaMemento, an adaptive memory-enhanced RL framework. Instead of just memorizing positive past experiences, we design a memory-reflection module that exploits both positive and negative experiences by learning to predict known local optimal policies based on real-time states. To effectively gather informative trajectories for the memory, we further introduce a fine-grained intrinsic motivation paradigm, where nuances in similar states can be precisely distinguished to guide exploration. The exploitation of past experiences and exploration of new policies are then adaptively coordinated by ensemble learning to approach the global optimum. Furthermore, we theoretically prove the superiority of our new intrinsic motivation and ensemble mechanism. From 59 quantitative and visualization experiments, we confirm that AdaMemento can distinguish subtle states for better exploration and effectively exploiting past experiences in memory, achieving significant improvement over previous methods. However, in sparse reward environments, policy updates become unstable and ineffective due to insufficient feedback (Bellemare et al., 2016; Liang et al., 2018). This significantly increases the difficulty of learning effective long-horizon policies. Memory offers a promising solution to the sparse reward problem, as humans can effectively learn from past experiences to avoid repeating mistakes in similar scenarios (Liu et al., 2021; Bransford & Johnson, 1972; Andrychowicz et al., 2017). Through memory, agents can utilize prior successful experiences to refine their policies in complex environments, hence reducing the reliance on dense reward feedback and improving both learning efficiency and policy stability (Pathak et al., 2017). Existing memory-based RL methods can be roughly categorized into two classes.

Model-based reinforcement learning has shown promise for improving sample efficiency and decision-making in complex environments. However, existing methods face challenges in training stability, robustness to noise, and computational efficiency. In this paper, we propose Bisimulation Metric for Model Predictive Control (BS-MPC), a novel approach that incorporates bisimulation metric loss in its objective function to directly optimize the encoder. This time-step-wise direct optimization enables the learned encoder to extract intrinsic information from the original state space while discarding irrelevant details and preventing the gradients and errors from diverging. BS-MPC improves training stability, robustness against input noise, and computational efficiency by reducing training time. We evaluate BS-MPC on both continuous control and image-based tasks from the DeepMind Control Suite, demonstrating superior performance and robustness compared to state-of-the-art baseline methods.

The problem of uncertainty is a feature of real world robotics problems and any control framework must contend with it in order to succeed in real applications tasks. Reinforcement Learning is no different, and epistemic uncertainty arising from model uncertainty or misspecification is a challenge well captured by the sim-to-real gap. A simple solution to this issue is domain randomization (DR), which unfortunately can result in conservative agents. As a remedy to this conservativeness, the use of universal policies that take additional information about the randomized domain has risen as an alternative solution, along with recurrent neural network-based controllers. Uncertainty-aware universal policies present a particularly compelling solution able to account for system identification uncertainties during deployment. In this paper, we reveal that the challenge of efficiently optimizing uncertainty-aware policies can be fundamentally reframed as solving the convex coverage set (CCS) problem within a multi-objective reinforcement learning (MORL) context. By introducing a novel Markov decision process (MDP) framework where each domain's performance is treated as an independent objective, we unify the training of uncertainty-aware policies with MORL approaches. This connection enables the application of MORL algorithms for domain randomization (DR), allowing for more efficient policy optimization. To illustrate this, we focus on the linear utility function, which aligns with the expectation in DR formulations, and propose a series of algorithms adapted from the MORL literature to solve the CCS, demonstrating their ability to enhance the performance of uncertainty-aware policies.

Controlling greenhouse crop production systems is a complex task due to uncertain and non-linear dynamics between crops, indoor and outdoor climate, and economics. The declining number of skilled growers necessitates the development of autonomous greenhouse control systems. Reinforcement Learning (RL) is a promising approach that can learn a control policy to automate greenhouse management. RL optimises a control policy through interactions with a model of the greenhouse while guided by an economic-based reward function. However, its application to real-world systems is limited due to discrepancies between models and real-world dynamics. Moreover, RL controllers may struggle to maintain state constraints while optimising the primary objective, especially when models inadequately capture the adverse effects of constraint violations on crop growth. Also, the generalisation to novel states, for example, due to unseen weather trajectories, is underexplored in RL-based greenhouse control. This work addresses these challenges through three key contributions. First, we present GreenLight-Gym, the first open-source environment designed for training and evaluating RL algorithms on the state-of-the-art greenhouse model GreenLight. GreenLight-Gym enables the community to benchmark RL-based control methodologies. Second, we compare two reward-shaping approaches, using either a multiplicative or additive penalty, to enforce state boundaries. The additive penalty achieves more stable training while better adhering to state constraints, while the multiplicative penalty yields marginally higher profits. Finally, we evaluate RL performance on a disjoint training and testing weather dataset, demonstrating improved generalisation to unseen conditions. Our environment and experiment scripts are open-sourced, facilitating innovative research on learning-based greenhouse control.

Robot soccer, in its full complexity, poses an unsolved research challenge. Current solutions heavily rely on engineered heuristic strategies, which lack robustness and adaptability. Deep reinforcement learning has gained significant traction in various complex robotics tasks such as locomotion, manipulation, and competitive games (e.g., AlphaZero, OpenAI Five), making it a promising solution to the robot soccer problem. This paper introduces MARLadona. A decentralized multi-agent reinforcement learning (MARL) training pipeline capable of producing agents with sophisticated team play behavior, bridging the shortcomings of heuristic methods. Further, we created an open-source multi-agent soccer environment based on Isaac Gym. Utilizing our MARL framework and a modified a global entity encoder as our core architecture, our approach achieves a 66.8% win rate against HELIOS agent, which employs a state-of-the-art heuristic strategy. Furthermore, we provided an in-depth analysis of the policy behavior and interpreted the agent's intention using the critic network.

Linear temporal logic (LTL) has recently been adopted as a powerful formalism for specifying complex, temporally extended tasks in reinforcement learning (RL). However, learning policies that efficiently satisfy arbitrary specifications not observed during training remains a challenging problem. Existing approaches suffer from several shortcomings: they are often only applicable to finite-horizon fragments of LTL, are restricted to suboptimal solutions, and do not adequately handle safety constraints. In this work, we propose a novel learning approach to address these concerns. Our method leverages the structure of Büchi automata, which explicitly represent the semantics of LTL specifications, to learn policies conditioned on sequences of truth assignments that lead to satisfying the desired formulae. Experiments in a variety of discrete and continuous domains demonstrate that our approach is able to zero-shot satisfy a wide range of finite-and infinite-horizon specifications, and outperforms existing methods in terms of both satisfaction probability and efficiency. One of the fundamental challenges in artificial intelligence (AI) is to create agents capable of following arbitrary instructions. While significant research efforts have been devoted to designing reinforcement learning (RL) agents that can complete tasks expressed in natural language (Oh et al., 2017; Goyal et al., 2019; Luketina et al., 2019), recent years have witnessed increased interest in formal languages to specify tasks in RL (Andreas et al., 2017; Camacho et al., 2019; Jothimurugan et al., 2021). Formal specification languages offer several desirable properties over natural language, such as well-defined semantics and compositionality, allowing for the specification of unambiguous, structured tasks (Vaezipoor et al., 2021; León et al., 2022). Recent works have furthermore shown that it is possible to automatically translate many natural language instructions into a relevant specification language, providing interpretable yet precise representations of tasks, which is especially important in safety-critical domains (León et al., 2021; Pan et al., 2023; Liu et al., 2023; Cohen et al., 2024). Linear temporal logic (LTL) (Pnueli, 1977) in particular has been adopted as a powerful formalism for instructing RL agents (Hasanbeig et al., 2018; Araki et al., 2021; Voloshin et al., 2023). LTL is an appealing specification language that allows for the definition of tasks in terms of high-level features of the environment.

Designing biological sequences with desired properties is a significant challenge due to the combinatorially vast search space and the high cost of evaluating each candidate sequence. To address these challenges, reinforcement learning (RL) methods, such as GFlowNets, utilize proxy models for rapid reward evaluation and annotated data for policy training. Although these approaches have shown promise in generating diverse and novel sequences, the limited training data relative to the vast search space often leads to the misspecification of proxy for out-of-distribution inputs. We introduce $\delta$-Conservative Search, a novel off-policy search method for training GFlowNets designed to improve robustness against proxy misspecification. The key idea is to incorporate conservativeness, controlled by parameter $\delta$, to constrain the search to reliable regions. Specifically, we inject noise into high-score offline sequences by randomly masking tokens with a Bernoulli distribution of parameter $\delta$ and then denoise masked tokens using the GFlowNet policy. Additionally, $\delta$ is adaptively adjusted based on the uncertainty of the proxy model for each data point. This enables the reflection of proxy uncertainty to determine the level of conservativeness. Experimental results demonstrate that our method consistently outperforms existing machine learning methods in discovering high-score sequences across diverse tasks-including DNA, RNA, protein, and peptide design-especially in large-scale scenarios.

On-policy reinforcement learning (RL) algorithms perform policy updates using i.i.d. trajectories collected by the current policy. However, after observing only a finite number of trajectories, on-policy sampling may produce data that fails to match the expected on-policy data distribution. This sampling error leads to noisy updates and data inefficient on-policy learning. Recent work in the policy evaluation setting has shown that non-i.i.d., off-policy sampling can produce data with lower sampling error than on-policy sampling can produce (Zhong et. al, 2022). Motivated by this observation, we introduce an adaptive, off-policy sampling method to improve the data efficiency of on-policy policy gradient algorithms. Our method, Proximal Robust On-Policy Sampling (PROPS), reduces sampling error by collecting data with a behavior policy that increases the probability of sampling actions that are under-sampled with respect to the current policy. We empirically evaluate PROPS on both continuous-action MuJoCo benchmark tasks as well discrete-action tasks and demonstrate that (1) PROPS decreases sampling error throughout training and (2) improves the data efficiency of on-policy policy gradient algorithms.

Navigating robots discreetly in human work environments while considering the possible privacy implications of robotic tasks presents significant challenges. Such scenarios are increasingly common, for instance, when robots transport sensitive objects that demand high levels of privacy in spaces crowded with human activities. While extensive research has been conducted on robotic path planning and social awareness, current robotic systems still lack the functionality of privacy-aware navigation in public environments. To address this, we propose a new framework for mobile robot navigation that leverages vision-language models to incorporate privacy awareness into adaptive path planning. Specifically, all potential paths from the starting point to the destination are generated using the A* algorithm. Concurrently, the vision-language model is used to infer the optimal path for privacy-awareness, given the environmental layout and the navigational instruction. This approach aims to minimize the robot's exposure to human activities and preserve the privacy of the robot and its surroundings. Experimental results on the S3DIS dataset demonstrate that our framework significantly enhances mobile robots' privacy awareness of navigation in human-shared public environments. Furthermore, we demonstrate the practical applicability of our framework by successfully navigating a robotic platform through real-world office environments. The supplementary video and code can be accessed via the following link: https://sites.google.com/view/privacy-aware-nav.

Deep reinforcement learning (DRL) has shown success in diverse domains such as robotics, computer games, and recommendation systems. However, like any other software system, DRL-based software systems are susceptible to faults that pose unique challenges for debugging and diagnosing. These faults often result in unexpected behavior without explicit failures and error messages, making debugging difficult and time-consuming. Therefore, automating the monitoring and diagnosis of DRL systems is crucial to alleviate the burden on developers. In this paper, we propose RLExplorer, the first fault diagnosis approach for DRL-based software systems. RLExplorer automatically monitors training traces and runs diagnosis routines based on properties of the DRL learning dynamics to detect the occurrence of DRL-specific faults. It then logs the results of these diagnoses as warnings that cover theoretical concepts, recommended practices, and potential solutions to the identified faults. We conducted two sets of evaluations to assess RLExplorer. Our first evaluation of faulty DRL samples from Stack Overflow revealed that our approach can effectively diagnose real faults in 83% of the cases. Our second evaluation of RLExplorer with 15 DRL experts/developers showed that (1) RLExplorer could identify 3.6 times more defects than manual debugging and (2) RLExplorer is easily integrated into DRL applications.