Large Language Models (LLMs) exhibit strong potential in mathematical reasoning, yet their effectiveness is often limited by a shortage of high-quality queries. This limitation necessitates scaling up computational responses through self-generated data, yet current methods struggle due to spurious correlated data caused by ineffective exploration across all reasoning stages. To address such challenge, we introduce **MARGE**: Improving **Ma**th **R**easoning with **G**uided **E**xploration, a novel method that enhances mathematical reasoning through hit-guided exploration. MARGE systematically explores intermediate reasoning states derived from self-generated solutions, enabling adequate exploration and improved credit assignment throughout the reasoning process. Notably, MARGE improves both single-shot accuracy and exploration diversity, mitigating a common trade-off in alignment methods. These results demonstrate MARGE's effectiveness in enhancing mathematical reasoning capabilities and unlocking the potential of scaling self-generated training data.

Visual representations play a crucial role in developing generalist robotic policies. Previous vision encoders, typically pre-trained with single-image reconstruction or two-image contrastive learning, tend to capture static information, often neglecting the dynamic aspects vital for embodied tasks. Recently, video diffusion models (VDMs) demonstrate the ability to predict future frames and showcase a strong understanding of physical world. We hypothesize that VDMs inherently produce visual representations that encompass both current static information and predicted future dynamics, thereby providing valuable guidance for robot action learning. Based on this hypothesis, we propose the Video Prediction Policy (VPP), which learns implicit inverse dynamics model conditioned on predicted future representations inside VDMs. To predict more precise future, we fine-tune pre-trained video foundation model on robot datasets along with internet human manipulation data. In experiments, VPP achieves a 18.6% relative improvement on the Calvin ABC-D generalization benchmark compared to the previous state-of-the-art, and demonstrates a 31.6% increase in success rates for complex real-world dexterous manipulation tasks. For your convenience, videos can be found at https://video-prediction-policy.github.io/

Recent advancements in Vision-Language-Action (VLA) models have leveraged pre-trained Vision-Language Models (VLMs) to improve the generalization capabilities. VLMs, typically pre-trained on vision-language understanding tasks, provide rich semantic knowledge and reasoning abilities. However, prior research has shown that VLMs often focus on high-level semantic content and neglect low-level features, limiting their ability to capture detailed spatial information and understand physical dynamics. These aspects, which are crucial for embodied control tasks, remain underexplored in existing pre-training paradigms. In this paper, we investigate the training paradigm for VLAs, and introduce UP-VLA, a Unified VLA model training with both multi-modal Understanding and future Prediction objectives, enhancing both high-level semantic comprehension and low-level spatial understanding. Experimental results show that UP-VLA achieves a 33% improvement on the Calvin ABC-D benchmark compared to the previous state-of-the-art method. Additionally, UP-VLA demonstrates improved success rates in real-world manipulation tasks, particularly those requiring precise spatial information.

Recent studies have successfully integrated large vision-language models (VLMs) into low-level robotic control by supervised fine-tuning (SFT) with expert robotic datasets, resulting in what we term vision-language-action (VLA) models. Although the VLA models are powerful, how to improve these large models during interaction with environments remains an open question. In this paper, we explore how to further improve these VLA models via Reinforcement Learning (RL), a commonly used fine-tuning technique for large models. However, we find that directly applying online RL to large VLA models presents significant challenges, including training instability that severely impacts the performance of large models, and computing burdens that exceed the capabilities of most local machines. To address these challenges, we propose iRe-VLA framework, which iterates between Reinforcement Learning and Supervised Learning to effectively improve VLA models, leveraging the exploratory benefits of RL while maintaining the stability of supervised learning. Experiments in two simulated benchmarks and a real-world manipulation suite validate the effectiveness of our method.

Evolutionary algorithms (EAs) have shown promise in solving sequential decision problems (SDPs) by simplifying them to static optimization problems and searching for the optimal policy parameters in a zeroth-order way. While these methods are highly versatile, they often suffer from high sample complexity due to their ignorance of the underlying temporal structures. In contrast, reinforcement learning (RL) methods typically formulate SDPs as Markov Decision Process (MDP). Although more sample efficient than EAs, RL methods are restricted to differentiable policies and prone to getting stuck in local optima. To address these issues, we propose a novel evolutionary framework Zeroth-Order Actor-Critic (ZOAC). We propose to use step-wise exploration in parameter space and theoretically derive the zeroth-order policy gradient. We further utilize the actor-critic architecture to effectively leverage the Markov property of SDPs and reduce the variance of gradient estimators. In each iteration, ZOAC employs samplers to collect trajectories with parameter space exploration, and alternates between first-order policy evaluation (PEV) and zeroth-order policy improvement (PIM). To evaluate the effectiveness of ZOAC, we apply it to a challenging multi-lane driving task, optimizing the parameters in a rule-based, non-differentiable driving policy that consists of three sub-modules: behavior selection, path planning, and trajectory tracking. We also compare it with gradient-based RL methods on three Gymnasium tasks, optimizing neural network policies with thousands of parameters. Experimental results demonstrate the strong capability of ZOAC in solving SDPs. ZOAC significantly outperforms EAs that treat the problem as static optimization and matches the performance of gradient-based RL methods even without first-order information, in terms of total average return across all tasks.

We focus on learning the zero-constraint-violation safe policy in model-free reinforcement learning (RL). Existing model-free RL studies mostly use the posterior penalty to penalize dangerous actions, which means they must experience the danger to learn from the danger. Therefore, they cannot learn a zero-violation safe policy even after convergence. To handle this problem, we leverage the safety-oriented energy functions to learn zero-constraint-violation safe policies and propose the safe set actor-critic (SSAC) algorithm. The energy function is designed to increase rapidly for potentially dangerous actions, locating the safe set on the action space. Therefore, we can identify the dangerous actions prior to taking them and achieve zero-constraint violation. Our major contributions are twofold. First, we use the data-driven methods to learn the energy function, which releases the requirement of known dynamics. Second, we formulate a constrained RL problem to solve the zero-violation policies. We prove that our Lagrangian-based constrained RL solutions converge to the constrained optimal zero-violation policies theoretically. The proposed algorithm is evaluated on the complex simulation environments and a hardware-in-loop (HIL) experiment with a real autonomous vehicle controller. Experimental results suggest that the converged policies in all environments achieve zero-constraint violation and comparable performance with model-based baseline.

Large language models (LLMs) encode a vast amount of semantic knowledge and possess remarkable understanding and reasoning capabilities. Previous work has explored how to ground LLMs in robotic tasks to generate feasible and executable textual plans. However, low-level execution in the physical world may deviate from the high-level textual plan due to environmental perturbations or imperfect controller design. In this paper, we propose DoReMi, a novel language model grounding framework that enables immediate Detection and Recovery from Misalignments between plan and execution. Specifically, we leverage LLMs to play a dual role, aiding not only in high-level planning but also generating constraints that can indicate misalignment during execution. Then vision language models (VLMs) are utilized to detect constraint violations continuously. Our pipeline can monitor the low-level execution and enable timely recovery if certain plan-execution misalignment occurs. Experiments on various complex tasks including robot arms and humanoid robots demonstrate that our method can lead to higher task success rates and shorter task completion times.

Diffusion models have demonstrated remarkable capabilities in image generation tasks, including image editing and video creation, representing a good understanding of the physical world. On the other line, diffusion models have also shown promise in robotic control tasks by denoising actions, known as diffusion policy. Although the diffusion generative model and diffusion policy exhibit distinct capabilities—image prediction and robotic action, respectively—they technically follow similar denoising process. In robotic tasks, the ability to predict future images and generate actions is highly correlated since they share the same underlying dynamics of the physical world. Building on this insight, we introduce \textbf{PAD}, a novel visual policy learning framework that unifies image \textbf{P}rediction and robot \textbf{A}ction within a joint \textbf{D}enoising process. Specifically, PAD utilizes Diffusion Transformers (DiT) to seamlessly integrate images and robot states, enabling the simultaneous prediction of future images and robot actions. Additionally, PAD supports co-training on both robotic demonstrations and large-scale video datasets and can be easily extended to other robotic modalities, such as depth images. PAD outperforms previous methods, achieving a significant 38.9\% relative improvement on the full Metaworld benchmark, by utilizing a single text-conditioned visual policy within a data-efficient imitation learning setting. Furthermore, PAD demonstrates superior generalization to unseen tasks in real-world robot manipulation settings with 28.0\% success rate increase compared to the strongest baseline. Videos of PAD can be found at https://sites.google.com/view/pad-paper

Safe reinforcement learning (RL) that solves constraint-satisfactory policies provides a promising way to the broader safety-critical applications of RL in real-world problems such as robotics. Among all safe RL approaches, model-based methods reduce training time violations further due to their high sample efficiency. However, lacking safety robustness against the model uncertainties remains an issue in safe model-based RL, especially in training time safety. In this paper, we propose a distributional reachability certificate (DRC) and its Bellman equation to address model uncertainties and characterize robust persistently safe states. Furthermore, we build a safe RL framework to resolve constraints required by the DRC and its corresponding shield policy. We also devise a line search method to maintain safety and reach higher returns simultaneously while leveraging the shield policy. Comprehensive experiments on classical benchmarks such as constrained tracking and navigation indicate that the proposed algorithm achieves comparable returns with much fewer constraint violations during training. Our code is available at https://github.com/ManUtdMoon/Distributional-Reachability-Policy-Optimization.Note to Practitioners—Although it has been proven that RL can be applied in complex robotics control tasks, the training process of an RL control policy induces frequent failures because the agent needs to learn safety through constraint violations. This issue hinders the promotion of RL because a large amount of failure of robots is too expensive to afford. This paper aims to reduce the training-time violations of RL-based control methods, enabling RL to be leveraged in a broader application area. To achieve the goal, we first introduce a safety quantity describing the distribution of potential constraint violations in the long term. By imposing constraints on the quantile of the safety distribution, we can realize safety robust to the model uncertainty, which is necessary for real-world robot learning with environment uncertainty. Second, we further devise a shield policy aiming to minimize the constraint violation. The policy will intervene when the agent is about to violate state constraints, further enhancing exploration safety. Third, we implement a line search method to find an action pursuing near-optimal performance when fulfilling safety requirements strictly. Our experimental results indicate that the proposed algorithm reduces training-time violations significantly while maintaining competitive task performance. We make a step towards applying RL safely in real-world tasks. Our future work includes conducting physical verification on real robots to evaluate the algorithm and improving safety further by starting from an initially safe control policy that comes from domain knowledge.

Safety is essential for reinforcement learning (RL) applied in the real world. Adding chance constraints (or probabilistic constraints) is a suitable way to enhance RL safety under uncertainty. Existing chance-constrained RL methods, such as the penalty methods and the Lagrangian methods, either exhibit periodic oscillations or learn an overconservative or unsafe policy. In this article, we address these shortcomings by proposing a separated proportional-integral Lagrangian (SPIL) algorithm. We first review the constrained policy optimization process from a feedback control perspective, which regards the penalty weight as the control input and the safe probability as the control output. Based on this, the penalty method is formulated as a proportional controller, and the Lagrangian method is formulated as an integral controller. We then unify them and present a proportional-integral Lagrangian method to get both their merits with an integral separation technique to limit the integral value to a reasonable range. To accelerate training, the gradient of safe probability is computed in a model-based manner. The convergence of the overall algorithm is analyzed. We demonstrate that our method can reduce the oscillations and conservatism of RL policy in a car-following simulation. To prove its practicality, we also apply our method to a real-world mobile robot navigation task, where our robot successfully avoids a moving obstacle with highly uncertain or even aggressive behaviors.

Humans are capable of abstracting various tasks as different combinations of multiple attributes. This perspective of compositionality is vital for human rapid learning and adaption since previous experiences from related tasks can be combined to generalize across novel compositional settings. In this work, we aim to achieve zero-shot policy generalization of Reinforcement Learning (RL) agents by leveraging the task compositionality. Our proposed method is a meta-RL algorithm with disentangled task representation, explicitly encoding different aspects of the tasks. Policy generalization is then performed by inferring unseen compositional task representations via the obtained disentanglement without extra exploration. The evaluation is conducted on three simulated tasks and a challenging real-world robotic insertion task. Experimental results demonstrate that our proposed method achieves policy generalization to unseen compositional tasks in a zero-shot manner.

Optimized trade execution is to sell (or buy) a given amount of assets in a given time with the lowest possible trading cost. Recently, reinforcement learning (RL) has been applied to optimized trade execution to learn smarter policies from market data. However, we find that many existing RL methods exhibit considerable overfitting which prevents them from real deployment. In this paper, we provide an extensive study on the overfitting problem in optimized trade execution. First, we model the optimized trade execution as offline RL with dynamic context (ORDC), where the context represents market variables that cannot be influenced by the trading policy and are collected in an offline manner. Under this framework, we derive the generalization bound and find that the overfitting issue is caused by large context space and limited context samples in the offline setting. Accordingly, we propose to learn compact representations for context to address the overfitting problem, either by leveraging prior knowledge or in an end-to-end manner. To evaluate our algorithms, we also implement a carefully designed simulator based on historical limit order book (LOB) data to provide a high-fidelity benchmark for different algorithms. Our experiments on the high-fidelity simulator demonstrate that our algorithms can effectively alleviate overfitting and achieve better performance.

The Hamilton-Jacobi-Bellman (HJB) equation serves as the necessary and sufficient condition for the optimal solution to the continuous-time (CT) optimal control problem (OCP). Compared with the infinite-horizon HJB equation, the solving of the finite-horizon (FH) HJB equation has been a long-standing challenge, because the partial time derivative of the value function is involved as an additional unknown term. To address this problem, this study first-time bridges the link between the partial time derivative and the terminal-time utility function, and thus it facilitates the use of the policy iteration (PI) technique to solve the CT FH OCPs. Based on this key finding, the FH approximate dynamic programming (ADP) algorithm is proposed leveraging an actor-critic framework. It is shown that the algorithm exhibits important properties in terms of convergence and optimality. Rather importantly, with the use of multilayer neural networks (NNs) in the actor-critic architecture, the algorithm is suitable for CT FH OCPs toward more general nonlinear and complex systems. Finally, the effectiveness of the proposed algorithm is demonstrated by conducting a series of simulations on both a linear quadratic regulator (LQR) problem and a nonlinear vehicle tracking problem.

Recent Semi-Supervised Object Detection (SS-OD) methods are mainly based on self-training, i.e., generating hard pseudo-labels by a teacher model on unlabeled data as supervisory signals. Although they achieved certain success, the limited labeled data in semi-supervised learning scales up the challenges of object detection. We analyze the challenges these methods meet with the empirical experiment results. We find that the massive False Negative samples and inferior localization precision lack consideration. Besides, the large variance of object sizes and class imbalance (i.e., the extreme ratio between back-ground and object) hinder the performance of prior arts. Further, we overcome these challenges by introducing a novel approach, Scale-Equivalent Distillation (SED), which is a simple yet effective end-to-end knowledge distillation framework robust to large object size variance and class imbalance. SED has several appealing benefits compared to the previous works. (1) SED imposes a consistency regularization to handle the large scale variance problem. (2) SED alleviates the noise problem from the False Negative samples and inferior localization precision. (3) A re-weighting strategy can implicitly screen the potential foreground regions of the unlabeled data to reduce the effect of class imbalance. Extensive experiments show that SED consistently outperforms the recent state-of-the-art methods on different datasets with significant margins. For example, it surpasses the supervised counterpart by more than 10 mAP when using 5% and 10% labeled data on MS-COCO.

Partially Observable Markov Decision Process (POMDP) provides a principled and generic framework to model real world sequential decision making processes but yet remains unsolved, especially for high dimensional continuous space and unknown models. The main challenge lies in how to accurately obtain the belief state, which is the probability distribution over the unobservable environment states given historical information. Accurately calculating this belief state is a precondition for obtaining an optimal policy of POMDPs. Recent advances in deep learning techniques show great potential to learn good belief states. However, existing methods can only learn approximated distribution with limited flexibility. In this paper, we introduce the \textbf{F}l\textbf{O}w-based \textbf{R}ecurrent \textbf{BE}lief \textbf{S}tate model (FORBES), which incorporates normalizing flows into the variational inference to learn general continuous belief states for POMDPs. Furthermore, we show that the learned belief states can be plugged into downstream RL algorithms to improve performance. In experiments, we show that our methods successfully capture the complex belief states that enable multi-modal predictions as well as high quality reconstructions, and results on challenging visual-motor control tasks show that our method achieves superior performance and sample efficiency.

Transformer has achieved great successes in learning vision and language representation, which is general across various downstream tasks. In visual control, learning transferable state representation that can transfer between different control tasks is important to reduce the training sample size. However, porting Transformer to sample-efficient visual control remains a challenging and unsolved problem. To this end, we propose a novel Control Transformer (CtrlFormer), possessing many appealing benefits that prior arts do not have. Firstly, CtrlFormer jointly learns self-attention mechanisms between visual tokens and policy tokens among different control tasks, where multitask representation can be learned and transferred without catastrophic forgetting. Secondly, we carefully design a contrastive reinforcement learning paradigm to train CtrlFormer, enabling it to achieve high sample efficiency, which is important in control problems. For example, in the DMControl benchmark, unlike recent advanced methods that failed by producing a zero score in the “Cartpole” task after transfer learning with 100k samples, CtrlFormer can achieve a state-of-the-art score with only 100k samples while maintaining the performance of previous tasks. The code and models are released in our project homepage.

Constrained reinforcement learning (CRL) has gained significant interest recently, since safety constraints satisfaction is critical for real-world problems. However, existing CRL methods constraining discounted cumulative costs generally lack rigorous definition and guarantee of safety. In contrast, in the safe control research, safety is defined as persistently satisfying certain state constraints. Such persistent safety is possible only on a subset of the state space, called feasible set, where an optimal largest feasible set exists for a given environment. Recent studies incorporate feasible sets into CRL with energy-based methods such as control barrier function (CBF), safety index (SI), and leverage prior conservative estimations of feasible sets, which harms the performance of the learned policy. To deal with this problem, this paper proposes the reachability CRL (RCRL) method by using reachability analysis to establish the novel self-consistency condition and characterize the feasible sets. The feasible sets are represented by the safety value function, which is used as the constraint in CRL. We use the multi-time scale stochastic approximation theory to prove that the proposed algorithm converges to a local optimum, where the largest feasible set can be guaranteed. Empirical results on different benchmarks validate the learned feasible set, the policy performance, and constraint satisfaction of RCRL, compared to CRL and safe control baselines.

Robotic grasping for a diverse set of objects is essential in many robot manipulation tasks. One promising approach is to learn deep grasping models from large training datasets of object images and grasp labels. However, empirical grasping datasets are typically sparsely labeled (i.e., a small number of successful grasp labels**Labels refer to marking the image to indicate a successful robotic grasp. in each image). The data sparsity issue can lead to insufficient supervision and false-negative labels, and thus results in poor learning results. This paper proposes a Maximum Likelihood Grasp Sampling Loss (MLGSL) to tackle the data sparsity issue. The proposed method supposes that successful grasps are stochastically sampled from the predicted grasp distribution and maximizes the observing likelihood. MLGSL is utilized for training a fully convolutional network that generates thousands of grasps simultaneously. Training results suggest that models based on MLGSL can learn to grasp with datasets composing of 2 labels per image. Compared to previous works, which require training datasets of 16 labels per image, MLGSL is 8× more data-efficient. Meanwhile, physical robot experiments demonstrate an equivalent performance at a 90.7% grasp success rate on household objects. Codes and videos are available at [1].

Reinforcement learning shows great potential to solve complex contact-rich robot manipulation tasks. However, the safety of using RL in the real world is a crucial problem, since unexpected dangerous collisions might happen when the RL policy is imperfect during training or in unseen scenarios. In this paper, we propose a contact-safe reinforcement learning framework for contact-rich robot manipulation, which maintains safety in both the task space and joint space. When the RL policy causes unexpected collisions between the robot arm and the environment, our framework is able to immediately detect the collision and ensure the contact force to be small. Furthermore, the end-effector is enforced to perform contact-rich tasks compliantly, while keeping robust to external disturbances. We train the RL policy in simulation and transfer it to the real robot. Real world experiments on robot wiping tasks show that our method is able to keep the contact force small both in task space and joint space even when the policy is under unseen scenario with unexpected collision, while rejecting the disturbances on the main task.

One of the key challenges in deploying RL to real-world applications is to adapt to variations of unknown environment contexts, such as changing terrains in robotic tasks and fluctuated bandwidth in congestion control. Existing works on adaptation to unknown environment contexts either assume the contexts are the same for the whole episode or assume the context variables are Markovian. However, in many real-world applications, the environment context usually stays stable for a stochastic period and then changes in an abrupt and unpredictable manner within an episode, resulting in a segment structure, which existing works fail to address. To leverage the segment structure of piecewise stable context in real-world applications, in this paper, we propose a \textit{\textbf{Se}gmented \textbf{C}ontext \textbf{B}elief \textbf{A}ugmented \textbf{D}eep~(SeCBAD)} RL method. Our method can jointly infer the belief distribution over latent context with the posterior over segment length and perform more accurate belief context inference with observed data within the current context segment. The inferred belief context can be leveraged to augment the state, leading to a policy that can adapt to abrupt variations in context. We demonstrate empirically that SeCBAD can infer context segment length accurately and outperform existing methods on a toy grid world environment and Mujuco tasks with piecewise-stable context.

Being able to discover diverse useful skills without external reward functions is beneficial in reinforcement learning research. Previous unsupervised skill discovery approaches mainly train different skills in parallel. Although impressive results have been provided, we found that parallel training procedure can sometimes block exploration when the state visited by different skills overlap, which leads to poor state coverage and restricts the diversity of learned skills. In this paper, we take a deeper look into this phenomenon and propose a novel framework to address this issue, which we call Recurrent Skill Training (ReST). Instead of training all the skills in parallel, ReST trains different skills one after another recurrently, along with a state coverage based intrinsic reward. We conduct experiments on a number of challenging 2D navigation environments and robotic locomotion environments. Evaluation results show that our proposed approach outperforms previous parallel training approaches in terms of state coverage and skill diversity. Videos of the discovered skills are available at https://sites.google.com/view/neurips22-rest.

Adapting to the changes in transition dynamics is essential in robotic applications. By learning a conditional policy with a compact context, context-aware meta-reinforcement learning provides a flexible way to adjust behavior according to dynamics changes. However, in real-world applications, the agent may encounter complex dynamics changes. Multiple confounders can influence the transition dynamics, making it challenging to infer accurate context for decision-making. This paper addresses such a challenge by decomposed mutual information optimization (DOMINO) for context learning, which explicitly learns a disentangled context to maximize the mutual information between the context and historical trajectories while minimizing the state transition prediction error. Our theoretical analysis shows that DOMINO can overcome the underestimation of the mutual information caused by multi-confounded challenges via learning disentangled context and reduce the demand for the number of samples collected in various environments. Extensive experiments show that the context learned by DOMINO benefits both model-based and model-free reinforcement learning algorithms for dynamics generalization in terms of sample efficiency and performance in unseen environments.

Unlike popular modularized framework, end-to-end autonomous driving seeks to solve the perception, decision and control problems in an integrated way, which can be more adapting to new scenarios and easier to generalize at scale. However, existing end-to-end approaches are often lack of interpretability, and can only deal with simple driving tasks like lane keeping. In this article, we propose an interpretable deep reinforcement learning method for end-to-end autonomous driving, which is able to handle complex urban scenarios. A sequential latent environment model is introduced and learned jointly with the reinforcement learning process. With this latent model, a semantic birdeye mask can be generated, which is enforced to connect with certain intermediate properties in today’s modularized framework for the purpose of explaining the behaviors of learned policy. The latent space also significantly reduces the sample complexity of reinforcement learning. Comparison tests in a realistic driving simulator show that the performance of our method in urban scenarios with crowded surrounding vehicles dominates many baselines including DQN, DDPG, TD3 and SAC. Moreover, through masked outputs, the learned model is able to provide a better explanation of how the car reasons about the driving environment.

Autonomous vehicles need to handle various traffic conditions and make safe and efficient decisions and maneuvers. However, on the one hand, a single optimization/sampling-based motion planner cannot efficiently generate safe trajectories in real time, particularly when there are many interactive vehicles near by. On the other hand, end-to-end learning methods cannot assure the safety of the outcomes. To address this challenge, we propose a hierarchical behavior planning framework with a set of low-level safe controllers and a high-level reinforcement learning algorithm (H-CtRL) as a coordinator for the low-level controllers. Safety is guaranteed by the low-level optimization/sampling-based controllers, while the high-level reinforcement learning algorithm makes H-CtRL an adaptive and efficient behavior planner. To train and test our proposed algorithm, we built a simulator that can reproduce traffic scenes using real-world datasets. The proposed HCtRL is proved to be effective in various realistic simulation scenarios, with satisfying performance in terms of both safety and efficiency.

Motion planning under uncertainty is of significant importance for safety-critical systems such as autonomous vehicles. Such systems have to satisfy necessary constraints (e.g., collision avoidance) with potential uncertainties coming from either disturbed system dynamics or noisy sensor measurements. However, existing motion planning methods cannot efficiently find the robust optimal solutions under general nonlinear and non-convex settings. In this paper, we formulate such problem as chance-constrained Gaussian belief space planning and propose the constrained iterative Linear Quadratic Gaussian (CILQG) algorithm as a real-time solution. In this algorithm, we iteratively calculate a Gaussian approximation of the belief and transform the chance-constraints. We evaluate the effectiveness of our method in simulations of autonomous driving planning tasks with static and dynamic obstacles. Results show that CILQG can handle uncertainties more appropriately and has faster computation time than baseline methods.

Model information can be used to predict future trajectories, so it has huge potential to avoid dangerous regions when applying reinforcement learning (RL) on real-world tasks, like autonomous driving. However, existing studies mostly use model-free constrained RL, which causes inevitable constraint violations. This paper proposes a model-based feasibility enhancement technique of constrained RL, which enhances the feasibility of policy using generalized control barrier function (GCBF) defined on the distance to constraint boundary. By using the model information, the policy can be optimized safely without violating actual safety constraints, and the sample efficiency is increased. The infeasibility in solving the constrained policy gradient is handled by an adaptive coefficient mechanism. We evaluate the proposed method in both simulations and real vehicle experiments in a complex autonomous driving collision avoidance task. The proposed method achieves up to four times fewer constraint violations and converges 3.36 times faster than baseline constrained RL approaches.

Safety is essential for reinforcement learning (RL) applied in real-world tasks like autonomous driving. Imposing chance constraints (or probabilistic constraints) is a suitable way to enhance RL safety under model uncertainty. Existing chance constrained RL methods like the penalty methods and the Lagrangian methods either exhibit periodic oscillations or learn an over-conservative or unsafe policy. In this paper, we address these shortcomings by elegantly combining these two methods and propose a separated proportional-integral Lagrangian (SPIL) algorithm. We first rewrite penalty methods as optimizing safe probability according to the proportional value of constraint violation, and Lagrangian methods as optimizing according to the integral value of the violation. Then we propose to add up both the integral and proportion values to optimize the policy, with an integral separation technique to limit the integral value within a reasonable range. Besides, the gradient of policy is computed in a model-based paradigm to accelerate training. The proposed method is proved to reduce oscillations and conservatism while ensuring safety by a car-following experiment.

Model-based reinforcement learning aims to improve the sample efficiency of policy learning by modeling the dynamics of the environment. Recently, the latent dynamics model is further developed to enable fast planning in a compact space. It summarizes the high-dimensional experiences of an agent, which mimics the memory function of humans. Learning policies via imagination with the latent model shows great potential for solving complex tasks. However, only considering memories from the true experiences in the process of imagination could limit its advantages. Inspired by the memory prosthesis proposed by neuroscientists, we present a novel model-based reinforcement learning framework called Imagining with Derived Memory (IDM). It enables the agent to learn policy from enriched diverse imagination with prediction-reliability weight, thus improving sample efficiency and policy robustness. Experiments on various high-dimensional visual control tasks in the DMControl benchmark demonstrate that IDM outperforms previous state-of-the-art methods in terms of policy robustness and further improves the sample efficiency of the model-based method.

Current autonomous driving systems are composed of a perception system and a decision system. Both of them are divided into multiple subsystems built up with lots of human heuristics. An end-to-end approach might clean up the system and avoid huge efforts of human engineering, as well as obtain better performance with increasing data and computation resources. Compared to the decision system, the perception system is more suitable to be designed in an end-to-end framework, since it does not require online driving exploration. In this paper, we propose a novel end-to-end approach for autonomous driving perception. A latent space is introduced to capture all relevant features useful for perception, which is learned through sequential latent representation learning. The learned end-to-end perception model is able to solve the detection, tracking, localization and mapping problems altogether with only minimum human engineering efforts and without storing any maps online. The proposed method is evaluated in a realistic urban driving simulator, with both camera image and lidar point cloud as sensor inputs. The codes and videos of this work are available at our github repo†and project website‡.

Probabilistic vehicle trajectory prediction is essential for robust safety of autonomous driving. Current methods for long-term trajectory prediction cannot guarantee the physical feasibility of predicted distribution. Moreover, their models cannot adapt to the driving policy of the predicted target human driver. In this work, we propose to overcome these two shortcomings by a Bayesian recurrent neural network model consisting of Bayesian-neural-network-based policy model and known physical model of the scenario. Bayesian neural network can ensemble complicated output distribution, enabling rich family of trajectory distribution. The embedded physical model ensures feasibility of the distribution. Moreover, the adopted gradient-based training method allows direct optimization for better performance in long prediction horizon. Furthermore, a particle-filter-based parameter adaptation algorithm is designed to adapt the policy Bayesian neural network to the predicted target online. Effectiveness of the proposed methods is verified with a toy example with multi-modal stochastic feedback gain and naturalistic car following data.

The decision and planning system for autonomous driving in urban environments is hard to design. Most current methods manually design the driving policy, which can be expensive to develop and maintain at scale. Instead, with imitation learning we only need to collect data and the computer will learn and improve the driving policy automatically. However, existing imitation learning methods for autonomous driving are hardly performing well for complex urban scenarios. Moreover, the safety is not guaranteed when we use a deep neural network policy. In this paper, we proposed a framework to learn the driving policy in urban scenarios efficiently given offline connected driving data, with a safety controller incorporated to guarantee safety at test time. The experiments show that our method can achieve high performance in realistic simulations of urban driving scenarios.

Although autonomous driving techniques have achieved great improvements, challenges still exist in decision making for variety of different scenarios under uncertain and interactive environments. A good decision maker must satisfy the following requirements: (1) Be in a generic and unified form to cover as more scenarios as possible. (2) Be able to interact properly with other moving obstacles under the uncertainty of their motions. In this paper, the continuous decision making (CDM) framework is proposed to formulate different driving scenarios in a unified way, which encodes the high level decision making information into a continuous reference trajectory that can be naturally combined with a lower level trajectory planner. Within the framework, a maximum interaction defensive policy (MIDP) is proposed, which calculates the best action to interact with stochastic moving obstacles while guaranteeing safety. The method is applied to a ramp merging scenario and the stochastic behavior models of the surrounding vehicles are learned from the NGSIM dataset. Simulations are shown to visualize and analyze the results.

Deep reinforcement learning has achieved great progress recently in domains such as learning to play Atari games from raw pixel input. The model-free characteristics of reinforcement learning free us from hand-encoding complex policies. However, for real world tasks such as autonomous driving, there are some complex sequential decision making processes that contain distinct behaviors. Due to the delayed rewards and the averaged gradient, it is pretty difficult for a flat deep reinforcement learning algorithm to learn a good policy. In this paper, we design a hierarchical neural network policy and propose a hierarchical policy gradient method to train the network with the semi markov decision process (SMDP) temporal abstraction formulation. We apply this method to a traffic light passing scenario in autonomous driving, where the vehicle has two distinct behaviors (e.g., pass and stop) and its primitive actions (e.g., acceleration) should follow the corresponding behavior. We show via simulation that our method is able to select correct decision and acts appropriately when the traffic light turns yellow. On the contrary, the flat reinforcement learning algorithm is not able to achieve a good performance and exhibits a large variance. Furthermore, the trained neural network modules are reusable in the future to cover more scenarios.

To address the problem of incorrect labels in training data for deep learning, we propose a novel and simple training strategy, Iterative Cross Learning (ICL), that significantly improves the classification accuracy of neural networks with training data that has noisy labels. We randomly partition the noisy training data into multiple separate subsets, each of which is used to train an independent network. After these independent networks predict labels for the original data, if the labels agree, we update the label with the predicted result for that data point, but otherwise we update the label with a random label, a key to the success of our proposed method. The process is repeated, possibly with several stages, to gradually improve the performance. Testing our method on MNIST and CIFAR-10 with partially shuffled labels, ICL significantly improves the classification accuracy of existing methods when the data labels have noise, especially in heavy noise situations. Moreover, the proposed method doesn't require any change to the underlying neural networks' structure or loss function, so it can also easily be combined with other existing methods that address noisy labels, improving their performance.

Conventional layered planning architecture temporally partitions the spatiotemporal motion planning by the path and speed, which is not suitable for lane change and overtaking scenarios with moving obstacles. In this paper, we propose to spatially partition the motion planning by longitudinal and lateral motions along the rough reference path in the Frenét Frame, which makes it possible to create linearized safety constraints for each layer in a variety of on-road driving scenarios. A generic environmental representation methodology is proposed with three topological elements and corresponding longitudinal constraints to compose all driving scenarios mentioned in this paper according to the overlap between the potential path of the autonomous vehicle and predicted path of other road users. Planners combining A* search and quadratic programming (QP) are designed to plan both rough long-term longitudinal motions and short-term trajectories to exploit the advantages of both search-based and optimization-based methods. Limits of vehicle kinematics and dynamics are considered in the planners to handle extreme cases. Simulation results show that the proposed framework can plan collision-free motions with high driving quality under complicated scenarios and emergency situations.