Decision-making for urban autonomous driving is challenging due to the stochastic nature of interactive traffic participants and the complexity of road structures. Although reinforcement learning (RL)-based decision-making scheme is promising to handle urban driving scenarios, it suffers from low sample efficiency and poor adaptability. In this paper, we propose Scene-Rep Transformer to improve the RL decision-making capabilities with better scene representation encoding and sequential predictive latent distillation. Specifically, a multi-stage Transformer (MST) encoder is constructed to model not only the interaction awareness between the ego vehicle and its neighbors but also intention awareness between the agents and their candidate routes. A sequential latent Transformer (SLT) with self-supervised learning objectives is employed to distill the future predictive information into the latent scene representation, in order to reduce the exploration space and speed up training. The final decision-making module based on soft actor-critic (SAC) takes as input the refined latent scene representation from the Scene-Rep Transformer and outputs driving actions. The framework is validated in five challenging simulated urban scenarios with dense traffic, and its performance is manifested quantitatively by the substantial improvements in data efficiency and performance in terms of success rate, safety, and efficiency. The qualitative results reveal that our framework is able to extract the intentions of neighbor agents to help make decisions and deliver more diversified driving behaviors.

Offline reinforcement learning (RL), which aims to learn an optimal policy using a previously collected static dataset, is an important paradigm of RL. Standard RL methods often perform poorly in this regime due to the function approximation errors on out-of-distribution actions. While a variety of regularization methods have been proposed to mitigate this issue, they are often constrained by policy classes with limited expressiveness that can lead to highly suboptimal solutions. In this paper, we propose representing the policy as a diffusion model, a recent class of highly-expressive deep generative models. We introduce Diffusion Q-learning (Diffusion-QL) that utilizes a conditional diffusion model to represent the policy. In our approach, we learn an action-value function and we add a term maximizing action-values into the training loss of the conditional diffusion model, which results in a loss that seeks optimal actions that are near the behavior policy. We show the expressiveness of the diffusion model-based policy, and the coupling of the behavior cloning and policy improvement under the diffusion model both contribute to the outstanding performance of Diffusion-QL. We illustrate the superiority of our method compared to prior works in a simple 2D bandit example with a multimodal behavior policy. We then show that our method can achieve state-of-the-art performance on the majority of the D4RL benchmark tasks.

The human brain is wired to be able to learn new things--and in all kinds of different ways, from imitating others to watching online explainer videos. What if robots could do the same thing? It is a question that ACM Prize recipient Pieter Abbeel, professor at the University of California, Berkeley and director of the Berkeley Robot Learning Lab, has spent his career researching. Here, we speak with Abbeel about his work and about the techniques he has developed to make it easier to teach robots. Let's start with deep reinforcement learning and the method you developed called Trust Region Policy Optimization.

We propose a nonparametric additive model for estimating interpretable value functions in reinforcement learning. Learning effective adaptive clinical interventions that rely on digital phenotyping features is a major for concern medical practitioners. With respect to spine surgery, different post-operative recovery recommendations concerning patient mobilization can lead to significant variation in patient recovery. While reinforcement learning has achieved widespread success in domains such as games, recent methods heavily rely on black-box methods, such neural networks. Unfortunately, these methods hinder the ability of examining the contribution each feature makes in producing the final suggested decision. While such interpretations are easily provided in classical algorithms such as Least Squares Policy Iteration, basic linearity assumptions prevent learning higher-order flexible interactions between features. In this paper, we present a novel method that offers a flexible technique for estimating action-value functions without making explicit parametric assumptions regarding their additive functional form. This nonparametric estimation strategy relies on incorporating local kernel regression and basis expansion to obtain a sparse, additive representation of the action-value function. Under this approach, we are able to locally approximate the action-value function and retrieve the nonlinear, independent contribution of select features as well as joint feature pairs. We validate the proposed approach with a simulation study, and, in an application to spine disease, uncover recovery recommendations that are inline with related clinical knowledge.

Deep reinforcement learning (RL) excels in various control tasks, yet the absence of safety guarantees hampers its real-world applicability. In particular, explorations during learning usually results in safety violations, while the RL agent learns from those mistakes. On the other hand, safe control techniques ensure persistent safety satisfaction but demand strong priors on system dynamics, which is usually hard to obtain in practice. To address these problems, we present Safe Set Guided State-wise Constrained Policy Optimization (S-3PO), a pioneering algorithm generating state-wise safe optimal policies with zero training violations, i.e., learning without mistakes. S-3PO first employs a safety-oriented monitor with black-box dynamics to ensure safe exploration. It then enforces a unique cost for the RL agent to converge to optimal behaviors within safety constraints. S-3PO outperforms existing methods in high-dimensional robotics tasks, managing state-wise constraints with zero training violation. This innovation marks a significant stride towards real-world safe RL deployment.

With the rising popularity of autonomous navigation research, Formula Student (FS) events are introducing a Driverless Vehicle (DV) category to their event list. This paper presents the initial investigation into utilising Deep Reinforcement Learning (RL) for end-to-end control of an autonomous FS race car for these competitions. We train two state-of-the-art RL algorithms in simulation on tracks analogous to the full-scale design on a Turtlebot2 platform. The results demonstrate that our approach can successfully learn to race in simulation and then transfer to a real-world racetrack on the physical platform. Finally, we provide insights into the limitations of the presented approach and guidance into the future directions for applying RL toward full-scale autonomous FS racing.

Risk-sensitive reinforcement learning (RL) has garnered significant attention in recent years due to the growing interest in deploying RL agents in real-world scenarios. A critical aspect of risk awareness involves modeling highly rare risk events (rewards) that could potentially lead to catastrophic outcomes. These infrequent occurrences present a formidable challenge for data-driven methods aiming to capture such risky events accurately. While risk-aware RL techniques do exist, their level of risk aversion heavily relies on the precision of the state-action value function estimation when modeling these rare occurrences. Our work proposes to enhance the resilience of RL agents when faced with very rare and risky events by focusing on refining the predictions of the extreme values predicted by the state-action value function distribution. To achieve this, we formulate the extreme values of the state-action value function distribution as parameterized distributions, drawing inspiration from the principles of extreme value theory (EVT). This approach effectively addresses the issue of infrequent occurrence by leveraging EVT-based parameterization. Importantly, we theoretically demonstrate the advantages of employing these parameterized distributions in contrast to other risk-averse algorithms. Our evaluations show that the proposed method outperforms other risk averse RL algorithms on a diverse range of benchmark tasks, each encompassing distinct risk scenarios.

Perimeter Control (PC) strategies have been proposed to address urban road network control in oversaturated situations by monitoring transfer flows of the Protected Network (PN). The uniform metering rate for cordon signals in existing studies ignores the variety of local traffic states at the intersection level, which may cause severe local traffic congestion and ruin the network stability. This paper introduces a semi-model dependent Multi-Agent Reinforcement Learning (MARL) framework to conduct PC with heterogeneous cordon signal behaviors. The proposed strategy integrates the MARL-based signal control method with centralized feedback PC policy and is applied to cordon signals of the PN. It operates as a two-stage system, with the feedback PC strategy detecting the overall traffic state within the PN and then distributing local instructions to cordon signals controlled by agents in the MARL framework. Each cordon signal acts independently and differently, creating a slack and distributed PC for the PN. The combination of the model-free and model-based methods is achieved by reconstructing the action-value function of the local agents with PC feedback reward without violating the integrity of the local signal control policy learned from the RL training process. Through numerical tests with different demand patterns in a microscopic traffic environment, the proposed PC strategy (a) is shown robustness, scalability, and transferability, (b) outperforms state-of-the-art model-based PC strategies in increasing network throughput, reducing cordon queue and carbon emission.

The increasing trend in adopting electric vehicles (EVs) will significantly impact the residential electricity demand, which results in an increased risk of transformer overload in the distribution grid. To mitigate such risks, there are urgent needs to develop effective EV charging controllers. Currently, the majority of the EV charge controllers are based on a centralized approach for managing individual EVs or a group of EVs. In this paper, we introduce a decentralized Multi-agent Reinforcement Learning (MARL) charging framework that prioritizes the preservation of privacy for EV owners. We employ the Centralized Training Decentralized Execution-Deep Deterministic Policy Gradient (CTDE-DDPG) scheme, which provides valuable information to users during training while maintaining privacy during execution. Our results demonstrate that the CTDE framework improves the performance of the charging network by reducing the network costs. Moreover, we show that the Peak-to-Average Ratio (PAR) of the total demand is reduced, which, in turn, reduces the risk of transformer overload during the peak hours.

Robot control using reinforcement learning has become popular, but its learning process generally terminates halfway through an episode for safety and time-saving reasons. This study addresses the problem of the most popular exception handling that temporal-difference (TD) learning performs at such termination. That is, by forcibly assuming zero value after termination, unintentionally implicit underestimation or overestimation occurs, depending on the reward design in the normal states. When the episode is terminated due to task failure, the failure may be highly valued with the unintentional overestimation, and the wrong policy may be acquired. Although this problem can be avoided by paying attention to the reward design, it is essential in practical use of TD learning to review the exception handling at termination. This paper therefore proposes a method to intentionally underestimate the value after termination to avoid learning failures due to the unintentional overestimation. In addition, the degree of underestimation is adjusted according to the degree of stationarity at termination, thereby preventing excessive exploration due to the intentional underestimation. Simulations and real robot experiments showed that the proposed method can stably obtain the optimal policies for various tasks and reward designs. https://youtu.be/AxXr8uFOe7M