Reinforcement Learning Community Research Articles

Bi-DexHands: Towards Human-Level Bimanual Dexterous Manipulation.

Achieving human-level dexterity in robotics remains a critical open problem. Even simple dexterous manipulation tasks pose significant difficulties due to the high number of degrees of freedom and the need for cooperation among heterogeneous agents (e.g., finger joints). While some researchers have utilized reinforcement learning (RL) to control a single hand in manipulating objects, tasks that require coordinated bimanual cooperation are still under-explored due to the fewer suitable environments, which can result in difficulties and sub-optimal performance. To address these challenges, we introduce Bi-DexHands, a simulator with two dexterous hands featuring 20 bimanual manipulation tasks and thousands of target objects, designed to match various levels of human motor skills based on cognitive science research. We developed Bi-DexHands in Issac Gym, enabling highly efficient RL training at over 30,000 frames per second using a single NVIDIA RTX 3090. Based on Bi-DexHands, we present a comprehensive evaluation of popular RL algorithms in different settings, including single-agent/multi-agent RL, offline RL, multi-task RL, and meta RL. Our findings show that on-policy algorithms, such as PPO, can master simple manipulation tasks that correspond to those of 48-month-old babies, such as catching a flying object or opening a bottle. Furthermore, multi-agent RL can improve the ability to perform manipulations that require skilled bimanual cooperation, such as lifting a pot or stacking blocks. Despite achieving success in individual tasks, current RL algorithms struggle to learn multiple manipulation skills in most multi-task and few-shot learning scenarios. This highlights the need for further research and development within the RL community.

GrowSpace: A reinforcement learning environment for plant architecture

Plants adapt over time to their surrounding conditions. We argue that no current plant simulator addresses plant responses to an environmental stimulus over time through gamification and sequential decision making. We propose the GrowSpace simulator that is built within a reinforcement learning framework, providing a sandbox environment to test out hypotheses. GrowSpace currently focuses on the behavioral response of plants to light displacement, with the objective of controlling a plant’s shape by moving a light source. The back-end of the simulator is implemented using the Space Colonisation Algorithm. This simulator serves as a low-dimensional test bed to visualize plant growth and movement with respect to a light stimulus and to better understand a plant’s reactions to these conditions. GrowSpace provides an artificial modular environment that is simulated to the advancements of plant biological systems all while using a reinforcement learning framework. Tasks are crafted and tested with artificial intelligence algorithms to further understand the concept of plant growth control through light. We equally demonstrate that sequential control of plant branching is hard challenge for the plant science community as well as the reinforcement learning community.

A Unified Lyapunov Framework for Finite-Sample Analysis of Reinforcement Learning Algorithms

Reinforcement learning (RL) is a paradigm where an agent learns to accomplish tasks by interacting with the environment, similar to how humans learn. RL is therefore viewed as a promising approach to achieve artificial intelligence, as evidenced by the remarkable empirical successes. However, many RL algorithms are theoretically not well-understood, especially in the setting where function approximation and off-policy sampling are employed. My thesis [1] aims at developing thorough theoretical understanding to the performance of various RL algorithms through finite-sample analysis.Since most of the RL algorithms are essentially stochastic approximation (SA) algorithms for solving variants of the Bellman equation, the first part of thesis is dedicated to the analysis of general SA involving a contraction operator, and under Markovian noise. We develop a Lyapunov approach where we construct a novel Lyapunov function called the generaled Moreau envelope. The results on SA enable us to establish finite-sample bounds of various RL algorithms in the tabular setting (cf. Part II of the thesis) and when using function approximation (cf. Part III of the thesis), which in turn provide theoretical insights to several important problems in the RL community, such as the efficiency of bootstrapping, the bias-variance trade-off in off-policy learning, and the stability of off-policy control.The main body of this document provides an overview of the contributions of my thesis.

Expected scalarised returns dominance: a new solution concept for multi-objective decision making

AbstractIn many real-world scenarios, the utility of a user is derived from a single execution of a policy. In this case, to apply multi-objective reinforcement learning, the expected utility of the returns must be optimised. Various scenarios exist where a user’s preferences over objectives (also known as the utility function) are unknown or difficult to specify. In such scenarios, a set of optimal policies must be learned. However, settings where the expected utility must be maximised have been largely overlooked by the multi-objective reinforcement learning community and, as a consequence, a set of optimal solutions has yet to be defined. In this work, we propose first-order stochastic dominance as a criterion to build solution sets to maximise expected utility. We also define a new dominance criterion, known as expected scalarised returns (ESR) dominance, that extends first-order stochastic dominance to allow a set of optimal policies to be learned in practice. Additionally, we define a new solution concept called the ESR set, which is a set of policies that are ESR dominant. Finally, we present a new multi-objective tabular distributional reinforcement learning (MOTDRL) algorithm to learn the ESR set in multi-objective multi-armed bandit settings.

ORSuite

Reinforcement learning (RL) has received widespread attention across multiple communities, but the experiments have focused primarily on large-scale game playing and robotics tasks. In this paper we introduce ORSuite, an open-source library containing environments, algorithms, and instrumentation for operational problems. Our package is designed to motivate researchers in the reinforcement learning community to develop and evaluate algorithms on operational tasks, and to consider the true multi-objective nature of these problems by considering metrics beyond cumulative reward.

Safe Reinforcement Learning Using Robust MPC

Reinforcement Learning (RL) has recently impressed the world with stunning results in various applications. While the potential of RL is now well-established, many critical aspects still need to be tackled, including safety and stability issues. These issues, while partially neglected by the RL community, are central to the control community which has been widely investigating them. Model Predictive Control (MPC) is one of the most successful control techniques because, among others, of its ability to provide such guarantees even for uncertain constrained systems. Since MPC is an optimization-based technique, optimality has also often been claimed. Unfortunately, the performance of MPC is highly dependent on the accuracy of the model used for predictions. In this paper, we propose to combine RL and MPC in order to exploit the advantages of both and, therefore, obtain a controller which is optimal and safe. We illustrate the results with a numerical example in simulations.

Unsupervised Grounding of Plannable First-Order Logic Representation from Images

Recently, there is an increasing interest in obtaining the relational structures of the environment in the Reinforcement Learning community. However, the resulting “relations” are not the discrete, logical predicates compatible with the symbolic reasoning such as classical planning or goal recognition. Meanwhile, Latplan (Asai and Fukunaga 2018) bridged the gap between deep-learning perceptual systems and symbolic classical planners. One key component of the system is a Neural Network called State AutoEncoder (SAE), which encodes an image-based input into a propositional representation compatible with classical planning. To get the best of both worlds, we propose First-Order State AutoEncoder, an unsupervised architecture for grounding the first-order logic predicates and facts. Each predicate models a relationship between objects by taking the interpretable arguments and returning a propositional value. In the experiment using 8Puzzle and a photo-realistic Blocksworld environment, we show that (1) the resulting predicates capture the interpretable relations (e.g., spatial), (2) they help to obtain the compact, abstract model of the environment, and finally, (3) the resulting model is compatible with symbolic classical planning.

Rainbow: Combining Improvements in Deep Reinforcement Learning

The deep reinforcement learning community has made several independent improvements to the DQN algorithm. However, it is unclear which of these extensions are complementary and can be fruitfully combined. This paper examines six extensions to the DQN algorithm and empirically studies their combination. Our experiments show that the combination provides state-of-the-art performance on the Atari 2600 benchmark, both in terms of data efficiency and final performance. We also provide results from a detailed ablation study that shows the contribution of each component to overall performance.

Teaching a Machine to Read Maps With Deep Reinforcement Learning

The ability to use a 2D map to navigate a complex 3D environment is quite remarkable, and even difficult for many humans. Localization and navigation is also an important problem in domains such as robotics, and has recently become a focus of the deep reinforcement learning community. In this paper we teach a reinforcement learning agent to read a map in order to find the shortest way out of a random maze it has never seen before. Our system combines several state-of-the-art methods such as A3C and incorporates novel elements such as a recurrent localization cell. Our agent learns to localize itself based on 3D first person images and an approximate orientation angle. The agent generalizes well to bigger mazes, showing that it learned useful localization and navigation capabilities.

Energy and Uncertainty: Models and Algorithms for Complex Energy Systems

The problem of controlling energy systems (generation, transmission, storage, investment) introduces a number of optimization problems which need to be solved in the presence of different types of uncertainty. We highlight several of these applications, using a simple energy storage problem as a case application. Using this setting, we describe a modeling framework based around five fundamental dimensions which is more natural than the standard canonical form widely used in the reinforcement learning community. The framework focuses on finding the best policy, where we identify four fundamental classes of policies consisting of policy function approximations (PFAs), cost function approximations (CFAs), policies based on value function approximations (VFAs), and lookahead policies. This organization unifies a number of competing strategies under a common umbrella.

Focus of Attention in Reinforcement Learning

Classification-based reinforcement learning (RL) methods have recently been pro- posed as an alternative to the traditional value-function based methods. These methods use a classifier to represent a policy, where the input (features) to the classifier is the state and the output (class label) for that state is the desired action. The reinforcement-learning community knows that focusing on more important states can lead to improved performance. In this paper, we investigate the idea of focused learning in the context of classification-based RL. Specifically, we define a useful notation of state importance, which we use to prove rigorous bounds on policy loss. Furthermore, we show that a classification-based RL agent may behave arbitrarily poorly if it treats all states as equally important.