In recent years, Reinforcement Learning (RL) has been applied to real-world problems with increasing success. Such applications often require to put constraints on the agent's behavior. Existing algorithms for constrained RL (CRL) rely on gradient descent-ascent, but this approach comes with a caveat. While these algorithms are guaranteed to converge on average, they do not guarantee last-iterate convergence, i.e., the current policy of the agent may never converge to the optimal solution. In practice, it is often observed that the policy alternates between satisfying the constraints and maximizing the reward, rarely accomplishing both objectives simultaneously. Here, we address this problem by introducing Reinforcement Learning with Optimistic Ascent-Descent (ReLOAD), a principled CRL method with guaranteed last-iterate convergence. We demonstrate its empirical effectiveness on a wide variety of CRL problems including discrete MDPs and continuous control. In the process we establish a benchmark of challenging CRL problems.

Deploying robots in real-world domains, such as households and flexible manufacturing lines, requires the robots to be taskable on demand. Linear temporal logic (LTL) is a widely-used specification language with a compositional grammar that naturally induces commonalities across tasks. However, the majority of prior research on reinforcement learning with LTL specifications treats every new formula independently. We propose LTL-Transfer, a novel algorithm that enables subpolicy reuse across tasks by segmenting policies for training tasks into portable transition-centric skills capable of satisfying a wide array of unseen LTL specifications while respecting safety-critical constraints. Experiments in a Minecraft-inspired domain show that LTL-Transfer can satisfy over 90% of 500 unseen tasks after training on only 50 task specifications and never violating a safety constraint. We also deployed LTL-Transfer on a quadruped mobile manipulator in an analog household environment to demonstrate its ability to transfer to many fetch and delivery tasks in a zero-shot fashion.

Heating and cooling systems in buildings account for 31% of global energy use, much of which are regulated by Rule Based Controllers (RBCs) that neither maximise energy efficiency nor minimise emissions by interacting optimally with the grid. Control via Reinforcement Learning (RL) has been shown to significantly improve building energy efficiency, but existing solutions require access to building-specific simulators or data that cannot be expected for every building in the world. In response, we show it is possible to obtain emission-reducing policies without such knowledge a priori--a paradigm we call zero-shot building control. We combine ideas from system identification and model-based RL to create PEARL (Probabilistic Emission-Abating Reinforcement Learning) and show that a short period of active exploration is all that is required to build a performant model. In experiments across three varied building energy simulations, we show PEARL outperforms an existing RBC once, and popular RL baselines in all cases, reducing building emissions by as much as 31% whilst maintaining thermal comfort. Our source code is available online via https://enjeeneer.io/projects/pearl/

Humans manipulate various kinds of fluids in their everyday life: creating latte art, scooping floating objects from water, rolling an ice cream cone, etc. Using robots to augment or replace human labors in these daily settings remain as a challenging task due to the multifaceted complexities of fluids. Previous research in robotic fluid manipulation mostly consider fluids governed by an ideal, Newtonian model in simple task settings (e.g., pouring water into a container). To evaluate robot learning algorithms on understanding and interacting with such complex fluid systems, a comprehensive virtual platform with versatile simulation capabilities and well-established tasks is needed. In this work, we introduce FluidLab, a simulation environment with a diverse set of manipulation tasks involving complex fluid dynamics. These tasks address interactions between solid and fluid as well as among multiple fluids. At the heart of our platform is a fully differentiable physics simulator, FluidEngine, providing GPU-accelerated simulations and gradient calculations for various material types and their couplings, extending the scope of the existing differentiable simulation engines. We identify several challenges for fluid manipulation learning by evaluating a set of reinforcement learning and trajectory optimization methods on our platform. To address these challenges, we propose several domain-specific optimization schemes coupled with differentiable physics, which are empirically shown to be effective in tackling optimization problems featured by fluid system's non-convex and nonsmooth properties. Furthermore, we demonstrate reasonable sim-to-real transfer by deploying optimized trajectories in real-world settings. FluidLab is publicly available at: https://fluidlab2023.github.io. Imagine you are fishing on the lakeside. Your hat falls into the water and starts to float out of reach. In order to get it back, you use your hands to paddle the water gently, generating a current that pulls the the hat back into reach.

Multimodal demonstrations provide robots with an abundance of information to make sense of the world. However, such abundance may not always lead to good performance when it comes to learning sensorimotor control policies from human demonstrations. Extraneous data modalities can lead to state over-specification, where the state contains modalities that are not only useless for decision-making but also can change data distribution across environments. State over-specification leads to issues such as the learned policy not generalizing outside of the training data distribution. In this work, we propose Masked Imitation Learning (MIL) to address state over-specification by selectively using informative modalities. Specifically, we design a masked policy network with a binary mask to block certain modalities. We develop a bi-level optimization algorithm that learns this mask to accurately filter over-specified modalities. We demonstrate empirically that MIL outperforms baseline algorithms in simulated domains including MuJoCo and a robot arm environment using the Robomimic dataset, and effectively recovers the environment-invariant modalities on a multimodal dataset collected on a real robot. Our project website presents supplemental details and videos of our results at: https://tinyurl.com/masked-il

Offline imitation learning (IL) promises the ability to learn performant policies from pre-collected demonstrations without interactions with the environment. However, imitating behaviors fully offline typically requires numerous expert data. To tackle this issue, we study the setting where we have limited expert data and supplementary suboptimal data. In this case, a well-known issue is the distribution shift between the learned policy and the behavior policy that collects the offline data. Prior works mitigate this issue by regularizing the KL divergence between the stationary state-action distributions of the learned policy and the behavior policy. We argue that such constraints based on exact distribution matching can be overly conservative and hamper policy learning, especially when the imperfect offline data is highly suboptimal. To resolve this issue, we present RelaxDICE, which employs an asymmetrically-relaxed f-divergence for explicit support regularization. Specifically, instead of driving the learned policy to exactly match the behavior policy, we impose little penalty whenever the density ratio between their stationary state-action distributions is upper bounded by a constant. Note that such formulation leads to a nested min-max optimization problem, which causes instability in practice. RelaxDICE addresses this challenge by supporting a closed-form solution for the inner maximization problem. Extensive empirical study shows that our method significantly outperforms the best prior offline IL method in six standard continuous control environments with over 30% performance gain on average, across 22 settings where the imperfect dataset is highly suboptimal.

Achieving and maintaining cooperation between agents to accomplish a common objective is one of the central goals of Multi-Agent Reinforcement Learning (MARL). Nevertheless in many real-world scenarios, separately trained and specialized agents are deployed into a shared environment, or the environment requires multiple objectives to be achieved by different coexisting parties. These variations among specialties and objectives are likely to cause mixed motives that eventually result in a social dilemma where all the parties are at a loss. In order to resolve this issue, we propose the Incentive Q-Flow (IQ-Flow) algorithm, which modifies the system's reward setup with an incentive regulator agent such that the cooperative policy also corresponds to the self-interested policy for the agents. Unlike the existing methods that learn to incentivize self-interested agents, IQ-Flow does not make any assumptions about agents' policies or learning algorithms, which enables the generalization of the developed framework to a wider array of applications. IQ-Flow performs an offline evaluation of the optimality of the learned policies using the data provided by other agents to determine cooperative and self-interested policies. Next, IQ-Flow uses meta-gradient learning to estimate how policy evaluation changes according to given incentives and modifies the incentive such that the greedy policy for cooperative objective and self-interested objective yield the same actions. We present the operational characteristics of IQ-Flow in Iterated Matrix Games. We demonstrate that IQ-Flow outperforms the state-of-the-art incentive design algorithm in Escape Room and 2-Player Cleanup environments. We further demonstrate that the pretrained IQ-Flow mechanism significantly outperforms the performance of the shared reward setup in the 2-Player Cleanup environment.

Effective traffic control methods have great potential in alleviating network congestion. Existing literature generally focuses on a single control approach, while few studies have explored the effectiveness of integrated and coordinated control approaches. This study considers two representative control approaches: ramp metering for freeways and perimeter control for homogeneous urban roads, and we aim to develop a deep reinforcement learning (DRL)-based coordinated control framework for large-scale networks. The main challenges are 1) there is a lack of efficient dynamic models for both freeways and urban roads; 2) the standard DRL method becomes ineffective due to the complex and non-stationary network dynamics. In view of this, we propose a novel meso-macro dynamic network model and first time develop a demonstration-guided DRL method to achieve large-scale coordinated ramp metering and perimeter control. The dynamic network model hybridizes the link and generalized bathtub models to depict the traffic dynamics of freeways and urban roads, respectively. For the DRL method, we incorporate demonstration to guide the DRL method for better convergence by introducing the concept of "teacher" and "student" models. The teacher models are traditional controllers (e.g., ALINEA, Gating), which provide control demonstrations. The student models are DRL methods, which learn from the teacher and aim to surpass the teacher's performance. To validate the proposed framework, we conduct two case studies in a small-scale network and a real-world large-scale traffic network in Hong Kong. The research outcome reveals the great potential of combining traditional controllers with DRL for coordinated control in large-scale networks.

Generative flow networks (GFlowNets), as an emerging technique, can be used as an alternative to reinforcement learning for exploratory control tasks. GFlowNet aims to generate distribution proportional to the rewards over terminating states, and to sample different candidates in an active learning fashion. GFlowNets need to form a DAG and compute the flow matching loss by traversing the inflows and outflows of each node in the trajectory. No experiments have yet concluded that GFlowNets can be used to handle continuous tasks. In this paper, we propose generative continuous flow networks (CFlowNets) that can be applied to continuous control tasks. First, we present the theoretical formulation of CFlowNets. Then, a training framework for CFlowNets is proposed, including the action selection process, the flow approximation algorithm, and the continuous flow matching loss function. Afterward, we theoretically prove the error bound of the flow approximation. The error decreases rapidly as the number of flow samples increases. Finally, experimental results on continuous control tasks demonstrate the performance advantages of CFlowNets compared to many reinforcement learning methods, especially regarding exploration ability.

Uncertainty quantification has been extensively used as a means to achieve efficient directed exploration in Reinforcement Learning (RL). However, state-of-the-art methods for continuous actions still suffer from high sample complexity requirements. Indeed, they either completely lack strategies for propagating the epistemic uncertainty throughout the updates, or they mix it with aleatoric uncertainty while learning the full return distribution (e.g., distributional RL). In this paper, we propose Wasserstein Actor-Critic (WAC), an actor-critic architecture inspired by the recent Wasserstein Q-Learning (WQL) \citep{wql}, that employs approximate Q-posteriors to represent the epistemic uncertainty and Wasserstein barycenters for uncertainty propagation across the state-action space. WAC enforces exploration in a principled way by guiding the policy learning process with the optimization of an upper bound of the Q-value estimates. Furthermore, we study some peculiar issues that arise when using function approximation, coupled with the uncertainty estimation, and propose a regularized loss for the uncertainty estimation. Finally, we evaluate our algorithm on standard MujoCo tasks as well as suite of continuous-actions domains, where exploration is crucial, in comparison with state-of-the-art baselines.