Online model predictive control (MPC) for piecewise affine (PWA) systems requires the online solution to an optimization problem that implicitly optimizes over the switching sequence of PWA regions, for which the computational burden can be prohibitive. Alternatively, the computation can be moved offline using explicit MPC; however, the online memory requirements and the offline computation can then become excessive. In this work we propose a solution in between online and explicit MPC, addressing the above issues by partially dividing the computation between online and offline. To solve the underlying MPC problem, a policy, learned offline, specifies the sequence of PWA regions that the dynamics must follow, thus reducing the complexity of the remaining optimization problem that solves over only the continuous states and control inputs. We provide a condition, verifiable during learning, that guarantees feasibility of the learned policy's output, such that an optimal continuous control input can always be found online. Furthermore, a method for iteratively generating training data offline allows the feasible policy to be learned efficiently, reducing the offline computational burden. A numerical experiment demonstrates the effectiveness of the method compared to both online and explicit MPC.
Autonomous intelligent agents must bridge computational challenges at disparate levels of abstraction, from the low-level spaces of sensory input and motor commands to the high-level domain of abstract reasoning and planning. A key question in designing such agents is how best to instantiate the representational space that will interface between these two levels -- ideally without requiring supervision in the form of expensive data annotations. These objectives can be efficiently achieved by representing the world in terms of objects (grounded in perception and action). In this work, we present a novel, brain-inspired, deep-learning architecture that learns from pixels to interpret, control, and reason about its environment, using object-centric representations. We show the utility of our approach through tasks in synthetic environments that require a combination of (high-level) logical reasoning and (low-level) continuous control. Results show that the agent can learn emergent conditional behavioural reasoning, such as $(A \to B) \land (\neg A \to C)$, as well as logical composition $(A \to B) \land (A \to C) \vdash A \to (B \land C)$ and XOR operations, and successfully controls its environment to satisfy objectives deduced from these logical rules. The agent can adapt online to unexpected changes in its environment and is robust to mild violations of its world model, thanks to dynamic internal desired goal generation. While the present results are limited to synthetic settings (2D and 3D activated versions of dSprites), which fall short of real-world levels of complexity, the proposed architecture shows how to manipulate grounded object representations, as a key inductive bias for unsupervised learning, to enable behavioral reasoning.
Sample efficiency in Reinforcement Learning (RL) has traditionally been driven by algorithmic enhancements. In this work, we demonstrate that scaling can also lead to substantial improvements. We conduct a thorough investigation into the interplay of scaling model capacity and domain-specific RL enhancements. These empirical findings inform the design choices underlying our proposed BRO (Bigger, Regularized, Optimistic) algorithm. The key innovation behind BRO is that strong regularization allows for effective scaling of the critic networks, which, paired with optimistic exploration, leads to superior performance. BRO achieves state-of-the-art results, significantly outperforming the leading model-based and model-free algorithms across 40 complex tasks from the DeepMind Control, MetaWorld, and MyoSuite benchmarks. BRO is the first model-free algorithm to achieve near-optimal policies in the notoriously challenging Dog and Humanoid tasks.
Diffusion models have garnered widespread attention in Reinforcement Learning (RL) for their powerful expressiveness and multimodality. It has been verified that utilizing diffusion policies can significantly improve the performance of RL algorithms in continuous control tasks by overcoming the limitations of unimodal policies, such as Gaussian policies, and providing the agent with enhanced exploration capabilities. However, existing works mainly focus on the application of diffusion policies in offline RL, while their incorporation into online RL is less investigated. The training objective of the diffusion model, known as the variational lower bound, cannot be optimized directly in online RL due to the unavailability of 'good' actions. This leads to difficulties in conducting diffusion policy improvement. To overcome this, we propose a novel model-free diffusion-based online RL algorithm, Q-weighted Variational Policy Optimization (QVPO). Specifically, we introduce the Q-weighted variational loss, which can be proved to be a tight lower bound of the policy objective in online RL under certain conditions. To fulfill these conditions, the Q-weight transformation functions are introduced for general scenarios. Additionally, to further enhance the exploration capability of the diffusion policy, we design a special entropy regularization term. We also develop an efficient behavior policy to enhance sample efficiency by reducing the variance of the diffusion policy during online interactions. Consequently, the QVPO algorithm leverages the exploration capabilities and multimodality of diffusion policies, preventing the RL agent from converging to a sub-optimal policy. To verify the effectiveness of QVPO, we conduct comprehensive experiments on MuJoCo benchmarks. The final results demonstrate that QVPO achieves state-of-the-art performance on both cumulative reward and sample efficiency.
We introduce the Continuous Arcade Learning Environment (CALE), an extension of the well-known Arcade Learning Environment (ALE) [Bellemare et al., 2013]. The CALE uses the same underlying emulator of the Atari 2600 gaming system (Stella), but adds support for continuous actions. This enables the benchmarking and evaluation of continuous-control agents (such as PPO [Schulman et al., 2017] and SAC [Haarnoja et al., 2018]) and value-based agents (such as DQN [Mnih et al., 2015] and Rainbow [Hessel et al., 2018]) on the same environment suite. We provide a series of open questions and research directions that CALE enables, as well as initial baseline results using Soft Actor-Critic. CALE is available as part of the ALE athttps://github.com/Farama-Foundation/Arcade-Learning-Environment.
Adversarial Imitation Learning is traditionally framed as a two-player zero-sum game between a learner and an adversarially chosen cost function, and can therefore be thought of as the sequential generalization of a Generative Adversarial Network (GAN). A prominent example of this framework is Generative Adversarial Imitation Learning (GAIL). However, in recent years, diffusion models have emerged as a non-adversarial alternative to GANs that merely require training a score function via regression, yet produce generations of a higher quality. In response, we investigate how to lift insights from diffusion modeling to the sequential setting. We propose diffusing states and performing score-matching along diffused states to measure the discrepancy between the expert's and learner's states. Thus, our approach only requires training score functions to predict noises via standard regression, making it significantly easier and more stable to train than adversarial methods. Theoretically, we prove first- and second-order instance-dependent bounds with linear scaling in the horizon, proving that our approach avoids the compounding errors that stymie offline approaches to imitation learning. Empirically, we show our approach outperforms GAN-style imitation learning baselines across various continuous control problems, including complex tasks like controlling humanoids to walk, sit, and crawl.
Data augmentation creates new data points by transforming the original ones for a reinforcement learning (RL) agent to learn from, which has been shown to be effective for the objective of improving the data efficiency of RL for continuous control. Prior work towards this objective has been largely restricted to perturbation-based data augmentation where new data points are created by perturbing the original ones, which has been impressively effective for tasks where the RL agent observes control states as images with perturbations including random cropping, shifting, etc. This work focuses on state-based control, where the RL agent can directly observe raw kinematic and task features, and considers an alternative data augmentation applied to these features based on Euclidean symmetries under transformations like rotations. We show that the default state features used in exiting benchmark tasks that are based on joint configurations are not amenable to Euclidean transformations. We therefore advocate using state features based on configurations of the limbs (i.e., the rigid bodies connected by the joints) that instead provide rich augmented data under Euclidean transformations. With minimal hyperparameter tuning, we show this new Euclidean data augmentation strategy significantly improves both data efficiency and asymptotic performance of RL on a wide range of continuous control tasks.
Reward shaping is critical in reinforcement learning (RL), particularly for complex tasks where sparse rewards can hinder learning. However, choosing effective shaping rewards from a set of reward functions in a computationally efficient manner remains an open challenge. We propose Online Reward Selection and Policy Optimization (ORSO), a novel approach that frames the selection of shaping reward function as an online model selection problem. ORSO automatically identifies performant shaping reward functions without human intervention with provable regret guarantees. We demonstrate ORSO's effectiveness across various continuous control tasks. Compared to prior approaches, ORSO significantly reduces the amount of data required to evaluate a shaping reward function, resulting in superior data efficiency and a significant reduction in computational time (up to 8 times). ORSO consistently identifies high-quality reward functions outperforming prior methods by more than 50% and on average identifies policies as performant as the ones learned using manually engineered reward functions by domain experts.
Reinforcement learning (RL) is rapidly reaching and surpassing human-level control capabilities. However, state-of-the-art RL algorithms often require timesteps and reaction times significantly faster than human capabilities, which is impractical in real-world settings and typically necessitates specialized hardware. Such speeds are difficult to achieve in the real world and often requires specialized hardware. We introduce Sequence Reinforcement Learning (SRL), an RL algorithm designed to produce a sequence of actions for a given input state, enabling effective control at lower decision frequencies. SRL addresses the challenges of learning action sequences by employing both a model and an actor-critic architecture operating at different temporal scales. We propose a "temporal recall" mechanism, where the critic uses the model to estimate intermediate states between primitive actions, providing a learning signal for each individual action within the sequence. Once training is complete, the actor can generate action sequences independently of the model, achieving model-free control at a slower frequency. We evaluate SRL on a suite of continuous control tasks, demonstrating that it achieves performance comparable to state-of-the-art algorithms while significantly reducing actor sample complexity. To better assess performance across varying decision frequencies, we introduce the Frequency-Averaged Score (FAS) metric. Our results show that SRL significantly outperforms traditional RL algorithms in terms of FAS, making it particularly suitable for applications requiring variable decision frequencies. Additionally, we compare SRL with model-based online planning, showing that SRL achieves superior FAS while leveraging the same model during training that online planners use for planning.
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.
Multi-objective reinforcement learning (MORL) excels at handling rapidly changing preferences in tasks that involve multiple criteria, even for unseen preferences. However, previous dominating MORL methods typically generate a fixed policy set or preference-conditioned policy through multiple training iterations exclusively for sampled preference vectors, and cannot ensure the efficient discovery of the Pareto front. Furthermore, integrating preferences into the input of policy or value functions presents scalability challenges, in particular as the dimension of the state and preference space grow, which can complicate the learning process and hinder the algorithm's performance on more complex tasks. To address these issues, we propose a two-stage Pareto front discovery algorithm called Constrained MORL (C-MORL), which serves as a seamless bridge between constrained policy optimization and MORL. Concretely, a set of policies is trained in parallel in the initialization stage, with each optimized towards its individual preference over the multiple objectives. Then, to fill the remaining vacancies in the Pareto front, the constrained optimization steps are employed to maximize one objective while constraining the other objectives to exceed a predefined threshold. Empirically, compared to recent advancements in MORL methods, our algorithm achieves more consistent and superior performances in terms of hypervolume, expected utility, and sparsity on both discrete and continuous control tasks, especially with numerous objectives (up to nine objectives in our experiments).
Learning from previously collected data via behavioral cloning or offline reinforcement learning (RL) is a powerful recipe for scaling generalist agents by avoiding the need for expensive online learning. Despite strong generalization in some respects, agents are often remarkably brittle to minor visual variations in control-irrelevant factors such as the background or camera viewpoint. In this paper, we present theDeepMind Control Visual Benchmark (DMC-VB), a dataset collected in the DeepMind Control Suite to evaluate the robustness of offline RL agents for solving continuous control tasks from visual input in the presence of visual distractors. In contrast to prior works, our dataset (a) combines locomotion and navigation tasks of varying difficulties, (b) includes static and dynamic visual variations, (c) considers data generated by policies with different skill levels, (d) systematically returns pairs of state and pixel observation, (e) is an order of magnitude larger, and (f) includes tasks with hidden goals. Accompanying our dataset, we propose three benchmarks to evaluate representation learning methods for pretraining, and carry out experiments on several recently proposed methods. First, we find that pretrained representations do not help policy learning on DMC-VB, and we highlight a large representation gap between policies learned on pixel observations and on states. Second, we demonstrate when expert data is limited, policy learning can benefit from representations pretrained on (a) suboptimal data, and (b) tasks with stochastic hidden goals. Our dataset and benchmark code to train and evaluate agents are available at: https://github.com/google-deepmind/dmc_vision_benchmark.
Recent advances in deep reinforcement learning have achieved impressive results in a wide range of complex tasks, but poor sample efficiency remains a major obstacle to real-world deployment. Soft actor-critic (SAC) mitigates this problem by combining stochastic policy optimization and off-policy learning, but its applicability is restricted to distributions whose gradients can be computed through the reparameterization trick. This limitation excludes several important examples such as the beta distribution, which was shown to improve the convergence rate of actor-critic algorithms in high-dimensional continuous control problems thanks to its bounded support. To address this issue, we investigate the use of implicit reparameterization, a powerful technique that extends the class of reparameterizable distributions. In particular, we use implicit reparameterization gradients to train SAC with the beta policy on simulated robot locomotion environments and compare its performance with common baselines. Experimental results show that the beta policy is a viable alternative, as it outperforms the normal policy and is on par with the squashed normal policy, which is the go-to choice for SAC. The code is available at https://github.com/lucadellalib/sac-beta.
Generative Flow Networks (GFlowNets) aim to generate diverse trajectories from a distribution in which the final states of the trajectories are proportional to the reward, serving as a powerful alternative to reinforcement learning for exploratory control tasks. However, the individual-flow matching constraint in GFlowNets limits their applications for multi-agent systems, especially continuous joint-control problems. In this paper, we propose a novel Multi-Agent generative Continuous Flow Networks (MACFN) method to enable multiple agents to perform cooperative exploration for various compositional continuous objects. Technically, MACFN trains decentralized individual-flow-based policies in a centralized global-flow-based matching fashion. During centralized training, MACFN introduces a continuous flow decomposition network to deduce the flow contributions of each agent in the presence of only global rewards. Then agents can deliver actions solely based on their assigned local flow in a decentralized way, forming a joint policy distribution proportional to the rewards. To guarantee the expressiveness of continuous flow decomposition, we theoretically derive a consistency condition on the decomposition network. Experimental results demonstrate that the proposed method yields results superior to the state-of-the-art counterparts and better exploration capability. Our code is available at https://github.com/isluoshuang/MACFN.
Deep reinforcement learning (RL) is a powerful approach to complex decision making. However, one issue that limits its practical application is its brittleness, sometimes failing to train in the presence of small changes in the environment. Motivated by the success of zero-shot transfer-where pre-trained models perform well on related tasks-we consider the problem of selecting a good set of training tasks to maximize generalization performance across a range of tasks. Given the high cost of training, it is critical to select training tasks strategically, but not well understood how to do so. We hence introduce Model-Based Transfer Learning (MBTL), which layers on top of existing RL methods to effectively solve contextual RL problems. MBTL models the generalization performance in two parts: 1) the performance set point, modeled using Gaussian processes, and 2) performance loss (generalization gap), modeled as a linear function of contextual similarity. MBTL combines these two pieces of information within a Bayesian optimization (BO) framework to strategically select training tasks. We show theoretically that the method exhibits sublinear regret in the number of training tasks and discuss conditions to further tighten regret bounds. We experimentally validate our methods using urban traffic and standard continuous control benchmarks. The experimental results suggest that MBTL can achieve up to 43x improved sample efficiency compared with canonical independent training and multi-task training. Further experiments demonstrate the efficacy of BO and the insensitivity to the underlying RL algorithm and hyperparameters. This work lays the foundations for investigating explicit modeling of generalization, thereby enabling principled yet effective methods for contextual RL.
Conventional uncertainty-aware temporal difference (TD) learning often assumes a zero-mean Gaussian distribution for TD errors, leading to inaccurate error representations and compromised uncertainty estimation. We introduce a novel framework for generalized Gaussian error modeling in deep reinforcement learning to enhance the flexibility of error distribution modeling by incorporating additional higher-order moment, particularly kurtosis, thereby improving the estimation and mitigation of data-dependent aleatoric uncertainty. We examine the influence of the shape parameter of the generalized Gaussian distribution (GGD) on aleatoric uncertainty and provide a closed-form expression that demonstrates an inverse relationship between uncertainty and the shape parameter. Additionally, we propose a theoretically grounded weighting scheme to address epistemic uncertainty by fully leveraging the GGD. We refine batch inverse variance weighting with bias reduction and kurtosis considerations, enhancing robustness. Experiments with policy gradient algorithms demonstrate significant performance gains.
For on-policy reinforcement learning, discretizing action space for continuous control can easily express multiple modes and is straightforward to optimize. However, without considering the inherent ordering between the discrete atomic actions, the explosion in the number of discrete actions can possess undesired properties and induce a higher variance for the policy gradient estimator. In this paper, we introduce a straightforward architecture that addresses this issue by constraining the discrete policy to be unimodal using Poisson probability distributions. This unimodal architecture can better leverage the continuity in the underlying continuous action space using explicit unimodal probability distributions. We conduct extensive experiments to show that the discrete policy with the unimodal probability distribution provides significantly faster convergence and higher performance for on-policy reinforcement learning algorithms in challenging control tasks, especially in highly complex tasks such as Humanoid. We provide theoretical analysis on the variance of the policy gradient estimator, which suggests that our attentively designed unimodal discrete policy can retain a lower variance and yield a stable learning process.
Despite the successes and progress of deep reinforcement learning over the last decade, several challenges remain that hinder its broader application. Some fundamental aspects to improve include data efficiency, generalization capability, and ability to learn in sparse-reward environments, which often require human-designed dense rewards. Meta-learning has emerged as a promising approach to address these issues by optimizing components of the learning algorithm to meet desired characteristics. Additionally, a different line of work has extensively studied the use of intrinsic rewards to enhance the exploration capabilities of algorithms. This work investigates how meta-learning can improve the training signal received by RL agents. The focus is on meta-learning intrinsic rewards under a framework that doesn't rely on the use of meta-gradients. We analyze and compare this approach to the use of extrinsic rewards and a meta-learned advantage function. The developed algorithms are evaluated on distributions of continuous control tasks with both parametric and non-parametric variations, and with only sparse rewards accessible for the evaluation tasks.
Reinforcement learning (RL) has progressed substantially over the past decade, with much of this progress being driven by benchmarks. Many benchmarks are focused on video or board games, and a large number of robotics benchmarks lack diversity and real-world applicability. In this paper, we aim to simplify the process of applying reinforcement learning in the 3D simulation league of Robocup, a robotic football competition. To this end, we introduce a Robocup-based RL environment based on the open source rcssserver3d soccer server, simple pre-defined tasks, and integration with a popular RL library, Stable Baselines 3. Our environment enables the creation of high-dimensional continuous control tasks within a robotics football simulation. In each task, an RL agent controls a simulated Nao robot, and can interact with the ball or other agents. We open-source our environment and training code at https://github.com/Michael-Beukman/RobocupGym.
We introduce Proximal Policy Distillation (PPD), a novel policy distillation method that integrates student-driven distillation and Proximal Policy Optimization (PPO) to increase sample efficiency and to leverage the additional rewards that the student policy collects during distillation. To assess the efficacy of our method, we compare PPD with two common alternatives, student-distill and teacher-distill, over a wide range of reinforcement learning environments that include discrete actions and continuous control (ATARI, Mujoco, and Procgen). For each environment and method, we perform distillation to a set of target student neural networks that are smaller, identical (self-distillation), or larger than the teacher network. Our findings indicate that PPD improves sample efficiency and produces better student policies compared to typical policy distillation approaches. Moreover, PPD demonstrates greater robustness than alternative methods when distilling policies from imperfect demonstrations. The code for the paper is released as part of a new Python library built on top of stable-baselines3 to facilitate policy distillation: `sb3-distill'.