In real-world applications, a reinforcement learning (RL) agent should consider multiple objectives and adhere to safety guidelines. To address these considerations, we propose a constrained multi-objective RL algorithm named constrained multi-objective gradient aggregator (CoMOGA). In the field of multi-objective optimization, managing conflicts between the gradients of the multiple objectives is crucial to prevent policies from converging to local optima. It is also essential to efficiently handle safety constraints for stable training and constraint satisfaction. We address these challenges straightforwardly by treating the maximization of multiple objectives as a constrained optimization problem (COP), where the constraints are defined to improve the original objectives. Existing safety constraints are then integrated into the COP, and the policy is updated by solving the COP, which ensures the avoidance of gradient conflicts. Despite its simplicity, CoMOGA guarantees convergence to global optima in a tabular setting. Through various experiments, we have confirmed that preventing gradient conflicts is critical, and the proposed method achieves constraint satisfaction across all tasks.

Distributional reinforcement learning improves performance by capturing environmental stochasticity, but a comprehensive theoretical understanding of its effectiveness remains elusive. In addition, the intractable element of the infinite dimensionality of distributions has been overlooked. In this paper, we present a regret analysis of distributional reinforcement learning with general value function approximation in a finite episodic Markov decision process setting. We first introduce a key notion of Bellman unbiasedness which is essential for exactly learnable and provably efficient distributional updates in an online manner. Among all types of statistical functionals for representing infinite-dimensional return distributions, our theoretical results demonstrate that only moment functionals can exactly capture the statistical information. Secondly, we propose a provably efficient algorithm, SF-LSVI, that achieves a tight regret bound of $\tilde{O}(d_E H^{\frac{3}{2}}\sqrt{K})$ where $H$ is the horizon, $K$ is the number of episodes, and $d_E$ is the eluder dimension of a function class.

To design reward that align with human goals, Reinforcement Learning from Human Feedback (RLHF) has emerged as a prominent technique for learning reward functions from human preferences and optimizing models using reinforcement learning algorithms. However, existing RLHF methods often misinterpret trajectories as being generated by an optimal policy, causing inaccurate likelihood estimation and suboptimal learning. To address this, we propose Policy-labeled Preference Learning (PPL) within the Direct Preference Optimization (DPO) framework, which resolves these likelihood mismatch problems by modeling human preferences with regret, reflecting the efficiency of executed policies. Additionally, we introduce a contrastive KL regularization term derived from regret-based principles to enhance sequential contrastive learning. Experiments in high-dimensional continuous control environments demonstrate PPL's significant improvements in offline RLHF performance and its effectiveness in online settings.

As the complexity of tasks addressed through reinforcement learning (RL) increases, the definition of reward functions also has become highly complicated. We introduce an RL method aimed at simplifying the reward-shaping process through intuitive strategies. Initially, instead of a single reward function composed of various terms, we define multiple reward and cost functions within a constrained multi-objective RL (CMORL) framework. For tasks involving sequential complex movements, we segment the task into distinct stages and define multiple rewards and costs for each stage. Finally, we introduce a practical CMORL algorithm that maximizes objectives based on these rewards while satisfying constraints defined by the costs. The proposed method has been successfully demonstrated across a variety of acrobatic tasks in both simulation and real-world environments. Additionally, it has been shown to successfully perform tasks compared to existing RL and constrained RL algorithms. Our code is available at https://github.com/rllab-snu/Stage-Wise-CMORL.

This paper addresses the problem of multi-risk measure agnostic multi-armed bandits in heavy-tailed reward settings. We propose a framework that leverages novel deviation inequalities for the $1$-Wasserstein distance to construct confidence intervals for Lipschitz risk measures. The distributional LCB (DistLCB) algorithm is introduced, which achieves asymptotic optimality by deriving the first lower bounds for risk measure aware bandits with explicit sub-optimality gap dependencies. The DistLCB is further extended to multi-risk objectives, which enables Pareto-optimal solutions that consider multiple aspects of reward distributions. Additionally, we provide a regret analysis that includes both gap-dependent and gap-independent bounds for multi-risk settings. Experiments validate the effectiveness of the proposed methods in synthetic and real-world applications.

Failures in propulsion components, such as propellers, can critically affect flight safety; thus, early failure detection, preferably before flight, is essential. Traditional fault-diagnosis methods typically rely on additional sensors or operational data, which may not be available or practical in all situations. This study addresses these challenges by introducing motor-electric-signal-based fault diagnosis that is independent of airframe configuration and can detect faults, even when the aircraft is not in operation. However, difficulties arise owing to poor class variance in motor-electric-signal data and the challenge of obtaining fault data. To overcome these issues, a semi-supervised learning model based on a modified variational autoencoder-generative adversarial network (VAE-GAN) is proposed, which predicts faults using only normal motor-electric-signal data. Additionally, a new preprocessing method and patch-based ensemble inference technique are introduced to improve the poor class-variance characteristics of the data, thereby enhancing the prediction performance. This work demonstrates that propeller faults can be successfully diagnosed using motor-electric signals without the need for additional sensors or fault-data acquisition. • Motor-electric-signal-based fault diagnosis for propellers without extra sensors. • Modified VAE-GAN model successfully detects faults using only normal signal data. • Event matrix preprocessing and patch ensemble achieve 91.8% accuracy and 98.4% AUROC.

Training agents that are robust to environmental changes remains a significant challenge in deep reinforcement learning (RL). Unsupervised environment design (UED) has recently emerged to address this issue by generating a set of training environments tailored to the agent's capabilities. While prior works demonstrate that UED has the potential to learn a robust policy, their performance is constrained by the capabilities of the environment generation. To this end, we propose a novel UED algorithm, adversarial environment design via regret-guided diffusion models (ADD). The proposed method guides the diffusion-based environment generator with the regret of the agent to produce environments that the agent finds challenging but conducive to further improvement. By exploiting the representation power of diffusion models, ADD can directly generate adversarial environments while maintaining the diversity of training environments, enabling the agent to effectively learn a robust policy. Our experimental results demonstrate that the proposed method successfully generates an instructive curriculum of environments, outperforming UED baselines in zero-shot generalization across novel, out-of-distribution environments.

The field of risk-constrained reinforcement learning (RCRL) has been developed to effectively reduce the likelihood of worst-case scenarios by explicitly handling risk-measure-based constraints. However, the nonlinearity of risk measures makes it challenging to achieve convergence and optimality. To overcome the difficulties posed by the nonlinearity, we propose a spectral risk measure-constrained RL algorithm, spectral-risk-constrained policy optimization (SRCPO), a bilevel optimization approach that utilizes the duality of spectral risk measures. In the bilevel optimization structure, the outer problem involves optimizing dual variables derived from the risk measures, while the inner problem involves finding an optimal policy given these dual variables. The proposed method, to the best of our knowledge, is the first to guarantee convergence to an optimum in the tabular setting. Furthermore, the proposed method has been evaluated on continuous control tasks and showed the best performance among other RCRL algorithms satisfying the constraints. Our code is available at https://github.com/rllab-snu/Spectral-Risk-Constrained-RL.

In this paper, we propose a signed distance field (SDF)-based deep Q-learning framework for multi-object re-arrangement. Our method learns to rearrange objects with non-prehensile manipulation, e.g., pushing, in unstructured environments. To reliably estimate Q-values in various scenes, we train the Q-network using an SDF-based scene graph as the state-goal representation. To this end, we introduce SDFGCN, a scalable Q-network structure which can estimate Q-values from a set of SDF images satisfying permutation invariance by using graph convolutional networks. In contrast to grasping-based rearrangement methods that rely on the performance of grasp predictive models for perception and movement, our approach enables rearrangements on unseen objects, including hard-to-grasp objects. Moreover, our method does not require any expert demonstrations. We observe that SDFGCN is capable of unseen objects in challenging configurations, both in the simulation and the real world.

Satisfying safety constraints in reinforcement learning (RL) is an important issue, especially in real-world applications. Many studies have approached safe RL with the Lagrangian method, which introduces dual variables. However, applying a trained policy with the optimal dual variable to a new environment can be hazardous since the optimal value of the dual variable, which represents a level of safety, depends on the environmental setting. To this end, we propose a new framework, dual variable actor-critic (DVAC), that solves the safe RL problem by simultaneously training a single policy over different safety levels. We introduce a universal policy and universal Q-function, which have a dual variable as an argument. Then, we extend the soft actor-critic so that the universal policy is guaranteed to converge to the Pareto optimal policy sets. We evaluate the proposed method in simulation and real-world environments. The universal policy learned with the proposed method ranges from extremely safe to high performance according to the dual variables, and is nearly Pareto optimal compared to policies learned with the baseline methods. In addition, the agent is able to adapt to environments with unseen state distributions without additional training by identifying a suitable dual variable using the proposed method.

In safety-critical robotic tasks, potential failures must be reduced, and multiple constraints must be met, such as avoiding collisions, limiting energy consumption, and maintaining balance. Thus, applying safe reinforcement learning (RL) in such robotic tasks requires to handle multiple constraints and use risk-averse constraints rather than risk-neutral constraints. To this end, we propose a trust region-based safe RL algorithm for multiple constraints called a safe distributional actor-critic (SDAC). Our main contributions are as follows: 1) introducing a gradient integration method to manage infeasibility issues in multi-constrained problems, ensuring theoretical convergence, and 2) developing a TD($\lambda$) target distribution to estimate risk-averse constraints with low biases. We evaluate SDAC through extensive experiments involving multi- and single-constrained robotic tasks. While maintaining high scores, SDAC shows 1.93 times fewer steps to satisfy all constraints in multi-constrained tasks and 1.78 times fewer constraint violations in single-constrained tasks compared to safe RL baselines. Code is available at: https://github.com/rllab-snu/Safe-Distributional-Actor-Critic.

Satisfying safety constraints is the top priority in safe reinforcement learning (RL). However, without proper exploration, an overly conservative policy such as freezing at the same position can be generated. To this end, we utilize maximum entropy RL methods for exploration. In particular, an RL method with Tsallis entropy maximization, called Tsallis actor-critic (TAC), is used to synthesize policies which can explore with more promising actions. In this paper, we propose a Tsallis entropy-regularized safe RL method for safer exploration, called SafeTAC. For more expressiveness, we extend the TAC to use a Gaussian mixture model policy, which improves the safety performance. To stabilize the training process, the retrace estimators for safety critics are formulated, and a safe policy update rule using a trust region method is proposed.

In reinforcement learning (RL), exploration is essential to achieve a globally optimal policy but unconstrained exploration can cause damages to robots and nearby people. To handle this safety issue in exploration, safe RL has been proposed to keep the agent under the specified safety constraints while maximizing cumulative rewards. This paper introduces a new safe RL method which can be applied to robots to operate under the safety constraints while learning. The key component of the proposed method is the safeguard module. The safeguard predicts the constraints in the near future and corrects actions such that the predicted constraints are not violated. Since actions are safely modified by the safeguard during exploration and policies are trained to imitate the corrected actions, the agent can safely explore. Additionally, the safeguard is sample efficient as it does not require long horizontal trajectories for training, so constraints can be satisfied within short time steps. The proposed method is extensively evaluated in simulation and experiments using a real robot. The results show that the proposed method achieves the best performance while satisfying safety constraints with minimal interaction with environments in all experiments.

We propose a novel autonomous driving frame-work that leverages graph-based features of roads, such as road positions and connections. The proposed method is divided into two parts: a low-level controller which follows the trajectory calculated by a graph-based path planner, and a high-level controller which determines the speed of the vehicle to follow the traffic flow. The high-level controller uses a road graphical neural network (Road-GNN), which encodes a road graph into latent features to perceive the surrounding environment. We use a 3D driving simulator to test the performance of Road-GNN, which is implemented based on the satellite image data of 30 roundabout intersections. To show that the proposed method can be generalized to various road environments, the proposed method is tested using roundabouts which are different from the training set. In the experiment, the proposed method successfully trains the agent and drives an ego-vehicle through various roundabout environments. The results show that the graph-based method is effective for autonomous driving.

In this paper, we consider autonomous driving of a vehicle using imitation learning. Generative adversarial imitation learning (GAIL) is a widely used algorithm for imitation learning. This algorithm leverages positive demonstrations to imitate the behavior of an expert. In this paper, we propose a novel method, called mixed generative adversarial imitation learning (MixGAIL), which incorporates both of expert demonstrations and negative demonstrations, such as vehicle collisions. To this end, the proposed method utilizes an occupancy measure and a constraint function. The occupancy measure is used to follow expert demonstrations and provides a positive feedback. On the other hand, the constraint function is used for negative demonstrations to assert a negative feedback. Experimental results show that the proposed algorithm converges faster than the other baseline methods. Also, hardware experiments using a real-world RC car shows an outstanding performance and faster convergence compared with existing methods.

A fundamental challenge in automated reasoning about programming assignments at scale is clustering student submissions based on their underlying algorithms. State-of-the-art clustering techniques are sensitive to control structure variations, cannot cluster buggy solutions with similar correct solutions, and either require expensive pair-wise program analyses or training efforts. We propose a novel technique that can cluster small imperative programs based on their algorithmic essence: (A) how the input space is partitioned into equivalence classes and (B) how the problem is uniquely addressed within individual equivalence classes. We capture these algorithmic aspects as two quantitative semantic program features that are merged into a program's vector representation. Programs are then clustered using their vector representations. The computation of our first semantic feature leverages model counting to identify the number of inputs belonging to an input equivalence class. The computation of our second semantic feature abstracts the program's data flow by tracking the number of occurrences of a unique pair of consecutive values of a variable during its lifetime. The comprehensive evaluation of our tool SemCluster on benchmarks drawn from solutions to small programming assignments shows that SemCluster (1) generates far fewer clusters than other clustering techniques, (2) precisely identifies distinct solution strategies, and (3) boosts the performance of clustering-based program repair, all within a reasonable amount of time.

Causality inference, such as dynamic taint anslysis, has many applications (e.g., information leak detection). It determines whether an event e is causally dependent on a preceding event c during execution. We develop a new causality inference engine LDX. Given an execution, it spawns a slave execution, in which it mutates c and observes whether any change is induced at e. To preclude non-determinism, LDX couples the executions by sharing syscall outcomes. To handle path differences induced by the perturbation, we develop a novel on-the-fly execution alignment scheme that maintains a counter to reflect the progress of execution. The scheme relies on program analysis and compiler transformation. LDX can effectively detect information leak and security attacks with an average overhead of 6.08% while running the master and the slave concurrently on separate CPUs, much lower than existing systems that require instruction level monitoring. Furthermore, it has much better accuracy in causality inference.

This paper presents Apex, a system that can automatically generate explanations for programming assignment bugs, regarding where the bugs are and how the root causes led to the runtime failures. It works by comparing the passing execution of a correct implementation (provided by the instructor) and the failing execution of the buggy implementation (submitted by the student). The technique overcomes a number of technical challenges caused by syntactic and semantic differences of the two implementations. It collects the symbolic traces of the executions and matches assignment statements in the two execution traces by reasoning about symbolic equivalence. It then matches predicates by aligning the control dependences of the matched assignment statements, avoiding direct matching of path conditions which are usually quite different. Our evaluation shows that Apex is every effective for 205 buggy real world student submissions of 4 programming assignments, and a set of 15 programming assignment type of buggy programs collected from stackoverflow.com, precisely pinpointing the root causes and capturing the causality for 94.5% of them. The evaluation on a standard benchmark set with over 700 student bugs shows similar results. A user study in the classroom shows that Apex has substantially improved student productivity.

Execution comparison has many applications in debugging, malware analysis, software feature identification, and intrusion detection. Existing comparison techniques have various limitations. Some can only compare at the system event level and require executions to take the same input. Some require storing instruction traces that are very space-consuming and have difficulty dealing with non-determinism. In this paper, we propose a novel dual execution technique that allows on-the-fly comparison at the instruction level. Only differences between the executions are recorded. It allows executions to proceed in a coupled mode such that they share the same input sequence with the same timing, reducing nondeterminism. It also allows them to proceed in a decoupled mode such that the user can interact with each one differently. Decoupled executions can be recoupled to share the same future inputs and facilitate further comparison. We have implemented a prototype and applied it to identifying functional components for reuse, comparative debugging with new GDB primitives, and understanding real world regression failures. Our results show that dual execution is a critical enabling technique for execution comparison.

Locating, extracting, and reusing the implementation of a feature within an existing binary program is challenging. This paper proposes a novel algorithm to identify modular functions corresponding to such features and to provide usable interfaces for the extracted functions. We provide a way to represent a desired feature with two executions that both execute the feature but with different inputs. Instead of reverse engineering the interface of a function, we wrap the existing interface and provide a simpler and more intuitive interface for the function through concretization and redirection. Experiments show that our technique can be applied to extract varied features from several real world applications including a malicious application.

We develop a logging and replay technique for real concurrent execution on multiple cores. Our technique directly works on binaries and does not require any hardware or complex software infrastructure support. We focus on minimizing logging overhead as it only logs a subset of system calls and thread spawns. Replay is on a single core. During replay, our technique first tries to follow only the event order in the log. However, due to schedule differences, replay may fail. An exploration process is then triggered to search for a schedule that allows the replay to make progress. Exploration is performed within a window preceding the point of replay failure. During exploration, our technique first tries to reorder synchronized blocks. If that does not lead to progress, it further reorders shared variable accesses. The exploration is facilitated by a sophisticated caching mechanism. Our experiments on real world programs and real workload show that the proposed technique has very low logging overhead (2.6% on average) and fast schedule reconstruction.

The first products of the Global Space-based Inter-Calibration System (GSICS) include bias monitoring and calibration corrections for the thermal infrared (IR) channels of current meteorological sensors on geostationary satellites. These use the hyperspectral Infrared Atmospheric Sounding Interferometer (IASI) on the low Earth orbit (LEO) Metop satellite as a common cross-calibration reference. This paper describes the algorithm, which uses a weighted linear regression, to compare collocated radiances observed from each pair of geostationary-LEO instruments. The regression coefficients define the GSICS Correction, and their uncertainties provide quality indicators, ensuring traceability to the selected community reference, IASI. Examples are given for the Meteosat, GOES, MTSAT, Fengyun-2, and COMS imagers. Some channels of these instruments show biases that vary with time due to variations in the thermal environment, stray light, and optical contamination. These results demonstrate how inter-calibration can be a powerful tool to monitor and correct biases, and help diagnose their root causes.