We present a convolutional framework which significantly reduces the complexity and thus, the computational effort for distributed reinforcement learning control of dynamical systems governed by partial differential equations (PDEs). Exploiting translational invariances, the high-dimensional distributed control problem can be transformed into a multi-agent control problem with many identical, uncoupled agents. Furthermore, using the fact that information is transported with finite velocity in many cases, the dimension of the agents' environment can be drastically reduced using a convolution operation over the state space of the PDE. In this setting, the complexity can be flexibly adjusted via the kernel width or by using a stride greater than one. Moreover, scaling from smaller to larger systems -- or the transfer between different domains -- becomes a straightforward task requiring little effort. We demonstrate the performance of the proposed framework using several PDE examples with increasing complexity, where stabilization is achieved by training a low-dimensional deep deterministic policy gradient agent using minimal computing resources.
Existing learning approaches to dexterous manipulation use demonstrations or interactions with the environment to train black-box neural networks that provide little control over how the robot learns the skills or how it would perform post training. These approaches pose significant challenges when implemented on physical platforms given that, during initial stages of training, the robot's behavior could be erratic and potentially harmful to its own hardware, the environment, or any humans in the vicinity. A potential way to address these limitations is to add constraints during learning that restrict and guide the robot's behavior during training as well as roll outs. Inspired by the success of constrained approaches in other domains, we investigate the effects of adding position-based constraints to a 24-DOF robot hand learning to perform object relocation using Constrained Policy Optimization. We find that a simple geometric constraint can ensure the robot learns to move towards the object sooner than without constraints. Further, training with this constraint requires a similar number of samples as its unconstrained counterpart to master the skill. These findings shed light on how simple constraints can help robots achieve sensible and safe behavior quickly and ease concerns surrounding hardware deployment. We also investigate the effects of the strictness of these constraints and report findings that provide insights into how different degrees of strictness affect learning outcomes. Our code is available at https://github.com/GT-STAR-Lab/constrained-rl-dexterous-manipulation.
Cooperative multi-robot tasks can benefit from heterogeneity in the robots' physical and behavioral traits. In spite of this, traditional Multi-Agent Reinforcement Learning (MARL) frameworks lack the ability to explicitly accommodate policy heterogeneity, and typically constrain agents to share neural network parameters. This enforced homogeneity limits application in cases where the tasks benefit from heterogeneous behaviors. In this paper, we crystallize the role of heterogeneity in MARL policies. Towards this end, we introduce Heterogeneous Graph Neural Network Proximal Policy Optimization (HetGPPO), a paradigm for training heterogeneous MARL policies that leverages a Graph Neural Network for differentiable inter-agent communication. HetGPPO allows communicating agents to learn heterogeneous behaviors while enabling fully decentralized training in partially observable environments. We complement this with a taxonomical overview that exposes more heterogeneity classes than previously identified. To motivate the need for our model, we present a characterization of techniques that homogeneous models can leverage to emulate heterogeneous behavior, and show how this "apparent heterogeneity" is brittle in real-world conditions. Through simulations and real-world experiments, we show that: (i) when homogeneous methods fail due to strong heterogeneous requirements, HetGPPO succeeds, and, (ii) when homogeneous methods are able to learn apparently heterogeneous behaviors, HetGPPO achieves higher resilience to both training and deployment noise.
Industrial robots are widely used in various manufacturing environments due to their efficiency in doing repetitive tasks such as assembly or welding. A common problem for these applications is to reach a destination without colliding with obstacles or other robot arms. Commonly used sampling-based path planning approaches such as RRT require long computation times, especially in complex environments. Furthermore, the environment in which they are employed needs to be known beforehand. When utilizing the approaches in new environments, a tedious engineering effort in setting hyperparameters needs to be conducted, which is time- and cost-intensive. On the other hand, Deep Reinforcement Learning has shown remarkable results in dealing with unknown environments, generalizing new problem instances, and solving motion planning problems efficiently. On that account, this paper proposes a Deep-Reinforcement-Learning-based motion planner for robotic manipulators. We evaluated our model against state-of-the-art sampling-based planners in several experiments. The results show the superiority of our planner in terms of path length and execution time.
Testing Deep Learning (DL) systems is a complex task as they do not behave like traditional systems would, notably because of their stochastic nature. Nonetheless, being able to adapt existing testing techniques such as Mutation Testing (MT) to DL settings would greatly improve their potential verifiability. While some efforts have been made to extend MT to the Supervised Learning paradigm, little work has gone into extending it to Reinforcement Learning (RL) which is also an important component of the DL ecosystem but behaves very differently from SL. This paper builds on the existing approach of MT in order to propose a framework, RLMutation, for MT applied to RL. Notably, we use existing taxonomies of faults to build a set of mutation operators relevant to RL and use a simple heuristic to generate test cases for RL. This allows us to compare different mutation killing definitions based on existing approaches, as well as to analyze the behavior of the obtained mutation operators and their potential combinations called Higher Order Mutation(s) (HOM). We show that the design choice of the mutation killing definition can affect whether or not a mutation is killed as well as the generated test cases. Moreover, we found that even with a relatively small number of test cases and operators we manage to generate HOM with interesting properties which can enhance testing capability in RL systems.
We consider the adversarial online multi-task reinforcement learning setting, where in each of $K$ episodes the learner is given an unknown task taken from a finite set of $M$ unknown finite-horizon MDP models. The learner's objective is to minimize its regret with respect to the optimal policy for each task. We assume the MDPs in $\mathcal{M}$ are well-separated under a notion of $\lambda$-separability, and show that this notion generalizes many task-separability notions from previous works. We prove a minimax lower bound of $\Omega(K\sqrt{DSAH})$ on the regret of any learning algorithm and an instance-specific lower bound of $\Omega(\frac{K}{\lambda^2})$ in sample complexity for a class of uniformly-good cluster-then-learn algorithms. We use a novel construction called 2-JAO MDP for proving the instance-specific lower bound. The lower bounds are complemented with a polynomial time algorithm that obtains $\tilde{O}(\frac{K}{\lambda^2})$ sample complexity guarantee for the clustering phase and $\tilde{O}(\sqrt{MK})$ regret guarantee for the learning phase, indicating that the dependency on $K$ and $\frac{1}{\lambda^2}$ is tight.
We present Orbit, a unified and modular framework for robot learning powered by NVIDIA Isaac Sim. It offers a modular design to easily and efficiently create robotic environments with photo-realistic scenes and high-fidelity rigid and deformable body simulation. With Orbit, we provide a suite of benchmark tasks of varying difficulty -- from single-stage cabinet opening and cloth folding to multi-stage tasks such as room reorganization. To support working with diverse observations and action spaces, we include fixed-arm and mobile manipulators with different physically-based sensors and motion generators. Orbit allows training reinforcement learning policies and collecting large demonstration datasets from hand-crafted or expert solutions in a matter of minutes by leveraging GPU-based parallelization. In summary, we offer an open-sourced framework that readily comes with 16 robotic platforms, 4 sensor modalities, 10 motion generators, more than 20 benchmark tasks, and wrappers to 4 learning libraries. With this framework, we aim to support various research areas, including representation learning, reinforcement learning, imitation learning, and task and motion planning. We hope it helps establish interdisciplinary collaborations in these communities, and its modularity makes it easily extensible for more tasks and applications in the future.
Model based reinforcement learning has proven to be more sample efficient than model free methods. On the other hand, the construction of a dynamics model in model based reinforcement learning has increased complexity. Data processing tasks in radio astronomy are such situations where the original problem which is being solved by reinforcement learning itself is the creation of a model. Fortunately, many methods based on heuristics or signal processing do exist to perform the same tasks and we can leverage them to propose the best action to take, or in other words, to provide a `hint'. We propose to use `hints' generated by the environment as an aid to the reinforcement learning process mitigating the complexity of model construction. We modify the soft actor critic algorithm to use hints and use the alternating direction method of multipliers algorithm with inequality constraints to train the agent. Results in several environments show that we get the increased sample efficiency by using hints as compared to model free methods.
Developing a general algorithm that learns to solve tasks across a wide range of applications has been a fundamental challenge in artificial intelligence. Although current reinforcement learning algorithms can be readily applied to tasks similar to what they have been developed for, configuring them for new application domains requires significant human expertise and experimentation. We present DreamerV3, a general algorithm that outperforms specialized methods across over 150 diverse tasks, with a single configuration. Dreamer learns a model of the environment and improves its behavior by imagining future scenarios. Robustness techniques based on normalization, balancing, and transformations enable stable learning across domains. Applied out of the box, Dreamer is the first algorithm to collect diamonds in Minecraft from scratch without human data or curricula. This achievement has been posed as a significant challenge in artificial intelligence that requires exploring farsighted strategies from pixels and sparse rewards in an open world. Our work allows solving challenging control problems without extensive experimentation, making reinforcement learning broadly applicable.
This work proposes a novel model-free Reinforcement Learning (RL) agent that is able to learn how to complete an unknown task having access to only a part of the input observation. We take inspiration from the concepts of visual attention and active perception that are characteristic of humans and tried to apply them to our agent, creating a hard attention mechanism. In this mechanism, the model decides first which region of the input image it should look at, and only after that it has access to the pixels of that region. Current RL agents do not follow this principle and we have not seen these mechanisms applied to the same purpose as this work. In our architecture, we adapt an existing model called recurrent attention model (RAM) and combine it with the proximal policy optimization (PPO) algorithm. We investigate whether a model with these characteristics is capable of achieving similar performance to state-of-the-art model-free RL agents that access the full input observation. This analysis is made in two Atari games, Pong and SpaceInvaders, which have a discrete action space, and in CarRacing, which has a continuous action space. Besides assessing its performance, we also analyze the movement of the attention of our model and compare it with what would be an example of the human behavior. Even with such visual limitation, we show that our model matches the performance of PPO+LSTM in two of the three games tested.
Inspired by organisms evolving through cooperation and competition between different populations on Earth, we study the emergence of artificial collective intelligence through massive-agent reinforcement learning. To this end, We propose a new massive-agent reinforcement learning environment, Lux, where dynamic and massive agents in two teams scramble for limited resources and fight off the darkness. In Lux, we build our agents through the standard reinforcement learning algorithm in curriculum learning phases and leverage centralized control via a pixel-to-pixel policy network. As agents co-evolve through self-play, we observe several stages of intelligence, from the acquisition of atomic skills to the development of group strategies. Since these learned group strategies arise from individual decisions without an explicit coordination mechanism, we claim that artificial collective intelligence emerges from massive-agent cooperation and competition. We further analyze the emergence of various learned strategies through metrics and ablation studies, aiming to provide insights for reinforcement learning implementations in massive-agent environments.
In this paper we describe Robofriend, a robotic teddy bear for telling stories to young children. Robofriend adapts its behavior to keep the childrens' attention using reinforcement learning.
Many challenging reinforcement learning (RL) problems require designing a distribution of tasks that can be applied to train effective policies. This distribution of tasks can be specified by the curriculum. A curriculum is meant to improve the results of learning and accelerate it. We introduce Success Induced Task Prioritization (SITP), a framework for automatic curriculum learning, where a task sequence is created based on the success rate of each task. In this setting, each task is an algorithmically created environment instance with a unique configuration. The algorithm selects the order of tasks that provide the fastest learning for agents. The probability of selecting any of the tasks for the next stage of learning is determined by evaluating its performance score in previous stages. Experiments were carried out in the Partially Observable Grid Environment for Multiple Agents (POGEMA) and Procgen benchmark. We demonstrate that SITP matches or surpasses the results of other curriculum design methods. Our method can be implemented with handful of minor modifications to any standard RL framework and provides useful prioritization with minimal computational overhead.
Despite some successful applications of goal-driven navigation, existing deep reinforcement learning (DRL)-based approaches notoriously suffers from poor data efficiency issue. One of the reasons is that the goal information is decoupled from the perception module and directly introduced as a condition of decision-making, resulting in the goal-irrelevant features of the scene representation playing an adversary role during the learning process. In light of this, we present a novel Goal-guided Transformer-enabled reinforcement learning (GTRL) approach by considering the physical goal states as an input of the scene encoder for guiding the scene representation to couple with the goal information and realizing efficient autonomous navigation. More specifically, we propose a novel variant of the Vision Transformer as the backbone of the perception system, namely Goal-guided Transformer (GoT), and pre-train it with expert priors to boost the data efficiency. Subsequently, a reinforcement learning algorithm is instantiated for the decision-making system, taking the goal-oriented scene representation from the GoT as the input and generating decision commands. As a result, our approach motivates the scene representation to concentrate mainly on goal-relevant features, which substantially enhances the data efficiency of the DRL learning process, leading to superior navigation performance. Both simulation and real-world experimental results manifest the superiority of our approach in terms of data efficiency, performance, robustness, and sim-to-real generalization, compared with other state-of-the-art (SOTA) baselines. The demonstration video (https://www.youtube.com/watch?v=aqJCHcsj4w0) and the source code (https://github.com/OscarHuangWind/DRL-Transformer-SimtoReal-Navigation) are also provided.
The ability for an agent to continuously learn new skills without catastrophically forgetting existing knowledge is of critical importance for the development of generally intelligent agents. Most methods devised to address this problem depend heavily on well-defined task boundaries, and thus depend on human supervision. Our task-agnostic method, Self-Activating Neural Ensembles (SANE), uses a modular architecture designed to avoid catastrophic forgetting without making any such assumptions. At the beginning of each trajectory, a module in the SANE ensemble is activated to determine the agent's next policy. During training, new modules are created as needed and only activated modules are updated to ensure that unused modules remain unchanged. This system enables our method to retain and leverage old skills, while growing and learning new ones. We demonstrate our approach on visually rich procedurally generated environments.
Classical reinforcement learning (RL) techniques are generally concerned with the design of decision-making policies driven by the maximisation of the expected outcome. Nevertheless, this approach does not take into consideration the potential risk associated with the actions taken, which may be critical in certain applications. To address that issue, the present research work introduces a novel methodology based on distributional RL to derive sequential decision-making policies that are sensitive to the risk, the latter being modelled by the tail of the return probability distribution. The core idea is to replace the $Q$ function generally standing at the core of learning schemes in RL by another function taking into account both the expected return and the risk. Named the risk-based utility function $U$, it can be extracted from the random return distribution $Z$ naturally learnt by any distributional RL algorithm. This enables to span the complete potential trade-off between risk minimisation and expected return maximisation, in contrast to fully risk-averse methodologies. Fundamentally, this research yields a truly practical and accessible solution for learning risk-sensitive policies with minimal modification to the distributional RL algorithm, and with an emphasis on the interpretability of the resulting decision-making process.
Designing better deep networks and better reinforcement learning (RL) algorithms are both important for deep RL. This work focuses on the former. Previous methods build the network with several modules like CNN, LSTM and Attention. Recent methods combine the Transformer with these modules for better performance. However, it requires tedious optimization skills to train a network composed of mixed modules, making these methods inconvenient to be used in practice. In this paper, we propose to design \emph{pure Transformer-based networks} for deep RL, aiming at providing off-the-shelf backbones for both the online and offline settings. Specifically, the Transformer in Transformer (TIT) backbone is proposed, which cascades two Transformers in a very natural way: the inner one is used to process a single observation, while the outer one is responsible for processing the observation history; combining both is expected to extract spatial-temporal representations for good decision-making. Experiments show that TIT can achieve satisfactory performance in different settings consistently.
In this work we introduce reinforcement learning techniques for solving lexicographic multi-objective problems. These are problems that involve multiple reward signals, and where the goal is to learn a policy that maximises the first reward signal, and subject to this constraint also maximises the second reward signal, and so on. We present a family of both action-value and policy gradient algorithms that can be used to solve such problems, and prove that they converge to policies that are lexicographically optimal. We evaluate the scalability and performance of these algorithms empirically, demonstrating their practical applicability. As a more specific application, we show how our algorithms can be used to impose safety constraints on the behaviour of an agent, and compare their performance in this context with that of other constrained reinforcement learning algorithms.
KL-regularized reinforcement learning from expert demonstrations has proved successful in improving the sample efficiency of deep reinforcement learning algorithms, allowing them to be applied to challenging physical real-world tasks. However, we show that KL-regularized reinforcement learning with behavioral reference policies derived from expert demonstrations can suffer from pathological training dynamics that can lead to slow, unstable, and suboptimal online learning. We show empirically that the pathology occurs for commonly chosen behavioral policy classes and demonstrate its impact on sample efficiency and online policy performance. Finally, we show that the pathology can be remedied by non-parametric behavioral reference policies and that this allows KL-regularized reinforcement learning to significantly outperform state-of-the-art approaches on a variety of challenging locomotion and dexterous hand manipulation tasks.
Efficient exploration strategy is one of essential issues in cooperative multi-agent reinforcement learning (MARL) algorithms requiring complex coordination. In this study, we introduce a new exploration method with the strangeness that can be easily incorporated into any centralized training and decentralized execution (CTDE)-based MARL algorithms. The strangeness refers to the degree of unfamiliarity of the observations that an agent visits. In order to give the observation strangeness a global perspective, it is also augmented with the the degree of unfamiliarity of the visited entire state. The exploration bonus is obtained from the strangeness and the proposed exploration method is not much affected by stochastic transitions commonly observed in MARL tasks. To prevent a high exploration bonus from making the MARL training insensitive to extrinsic rewards, we also propose a separate action-value function trained by both extrinsic reward and exploration bonus, on which a behavioral policy to generate transitions is designed based. It makes the CTDE-based MARL algorithms more stable when they are used with an exploration method. Through a comparative evaluation in didactic examples and the StarCraft Multi-Agent Challenge, we show that the proposed exploration method achieves significant performance improvement in the CTDE-based MARL algorithms.