Domain Randomization (DR) is commonly used for sim2real transfer of reinforcement learning (RL) policies in robotics. Most DR approaches require a simulator with a fixed set of tunable parameters from the start of the training, from which the parameters are randomized simultaneously to train a robust model for use in the real world. However, the combined randomization of many parameters increases the task difficulty and might result in sub-optimal policies. To address this problem and to provide a more flexible training process, we propose Continual Domain Randomization (CDR) for RL that combines domain randomization with continual learning to enable sequential training in simulation on a subset of randomization parameters at a time. Starting from a model trained in a non-randomized simulation where the task is easier to solve, the model is trained on a sequence of randomizations, and continual learning is employed to remember the effects of previous randomizations. Our robotic reaching and grasping tasks experiments show that the model trained in this fashion learns effectively in simulation and performs robustly on the real robot while matching or outperforming baselines that employ combined randomization or sequential randomization without continual learning. Our code and videos are available at https://continual-dr.github.io/.

When writing and talking, people sometimes pause to think. Although reasoning-focused works have often framed reasoning as a method of answering questions or completing agentic tasks, reasoning is implicit in almost all written text. For example, this applies to the steps not stated between the lines of a proof or to the theory of mind underlying a conversation. In the Self-Taught Reasoner (STaR, Zelikman et al. 2022), useful thinking is learned by inferring rationales from few-shot examples in question-answering and learning from those that lead to a correct answer. This is a highly constrained setting -- ideally, a language model could instead learn to infer unstated rationales in arbitrary text. We present Quiet-STaR, a generalization of STaR in which LMs learn to generate rationales at each token to explain future text, improving their predictions. We address key challenges, including 1) the computational cost of generating continuations, 2) the fact that the LM does not initially know how to generate or use internal thoughts, and 3) the need to predict beyond individual next tokens. To resolve these, we propose a tokenwise parallel sampling algorithm, using learnable tokens indicating a thought's start and end, and an extended teacher-forcing technique. Encouragingly, generated rationales disproportionately help model difficult-to-predict tokens and improve the LM's ability to directly answer difficult questions. In particular, after continued pretraining of an LM on a corpus of internet text with Quiet-STaR, we find zero-shot improvements on GSM8K (5.9%$\rightarrow$10.9%) and CommonsenseQA (36.3%$\rightarrow$47.2%) and observe a perplexity improvement of difficult tokens in natural text. Crucially, these improvements require no fine-tuning on these tasks. Quiet-STaR marks a step towards LMs that can learn to reason in a more general and scalable way.

The prevailing approach to aligning Large Language Models (LLMs) typically relies on human or AI feedback and assumes access to specific types of preference datasets. In our work, we question the efficacy of such datasets and explore various scenarios where alignment with expert demonstrations proves more realistic. We build a sequential decision-making framework to formulate the problem of aligning LLMs using demonstration datasets. Drawing insights from inverse reinforcement learning and imitation learning, we introduce various approaches for divergence minimization in the LLM alignment tasks. Our analysis highlights the mass-covering and mode-seeking behaviors of these different approaches. Inclusively, we examine the pros and cons of the classical supervised fine-tuning method, elaborating on scenarios where different methods shine.

To solve complex tasks under resource constraints, reinforcement learning (RL) agents need to be simple, efficient, and scalable, addressing (1) large state spaces and (2) the continuous accumulation of interaction data. We propose HyperAgent, an RL framework featuring the hypermodel and index sampling schemes that enable computation-efficient incremental approximation for the posteriors associated with general value functions without the need for conjugacy, and data-efficient action selection. Implementing HyperAgent is straightforward, requiring only one additional module beyond what is necessary for Double-DQN. HyperAgent stands out as the first method to offer robust performance in large-scale deep RL benchmarks while achieving provably scalable per-step computational complexity and attaining sublinear regret under tabular assumptions. HyperAgent can solve Deep Sea hard exploration problems with episodes that optimally scale with problem size and exhibits significant efficiency gains in both data and computation under the Atari benchmark. The core of our theoretical analysis is the sequential posterior approximation argument, enabled by the first analytical tool for sequential random projection -- a non-trivial martingale extension of the Johnson-Lindenstrauss. This work bridges the theoretical and practical realms of RL, establishing a new benchmark for RL algorithm design.

Causality has become a fundamental approach for explaining the relationships between events, phenomena, and outcomes in various fields of study. It has invaded various fields and applications, such as medicine, healthcare, economics, finance, fraud detection, cybersecurity, education, public policy, recommender systems, anomaly detection, robotics, control, sociology, marketing, and advertising. In this paper, we survey its development over the past five decades, shedding light on the differences between causality and other approaches, as well as the preconditions for using it. Furthermore, the paper illustrates how causality interacts with new approaches such as Artificial Intelligence (AI), Generative AI (GAI), Machine and Deep Learning, Reinforcement Learning (RL), and Fuzzy Logic. We study the impact of causality on various fields, its contribution, and its interaction with state-of-the-art approaches. Additionally, the paper exemplifies the trustworthiness and explainability of causality models. We offer several ways to evaluate causality models and discuss future directions.

Motivated by agility, 3D mobility, and low-risk operation compared to human-operated management systems of autonomous unmanned aerial vehicles (UAVs), this work studies UAV-based active wildfire monitoring where a UAV detects fire incidents in remote areas and tracks the fire frontline. A UAV path planning solution is proposed considering realistic wildfire management missions, where a single low-altitude drone with limited power and flight time is available. Noting the limited field of view of commercial low-altitude UAVs, the problem formulates as a partially observable Markov decision process (POMDP), in which wildfire progression outside the field of view causes inaccurate state representation that prevents the UAV from finding the optimal path to track the fire front in limited time. Common deep reinforcement learning (DRL)-based trajectory planning solutions require diverse drone-recorded wildfire data to generalize pre-trained models to real-time systems, which is not currently available at a diverse and standard scale. To narrow down the gap caused by partial observability in the space of possible policies, a belief-based state representation with broad, extensive simulated data is proposed where the beliefs (i.e., ignition probabilities of different grid areas) are updated using a Bayesian framework for the cells within the field of view. The performance of the proposed solution in terms of the ratio of detected fire cells and monitored ignited area (MIA) is evaluated in a complex fire scenario with multiple rapidly growing fire batches, indicating that the belief state representation outperforms the observation state representation both in fire coverage and the distance to fire frontline.

We have developed a parallel wire-driven monopedal robot, RAMIEL, which has both speed and power due to the parallel wire mechanism and a long acceleration distance. RAMIEL is capable of jumping high and continuously, and so has high performance in traveling. On the other hand, one of the drawbacks of a minimal parallel wire-driven robot without joint encoders is that the current joint velocities estimated from the wire lengths oscillate due to the elongation of the wires, making the values unreliable. Therefore, despite its high performance, the control of the robot is unstable, and in 10 out of 16 jumps, the robot could only jump up to two times continuously. In this study, we propose a method to realize a continuous jumping motion by reinforcement learning in simulation, and its application to the actual robot. Because the joint velocities oscillate with the elongation of the wires, they are not used directly, but instead are inferred from the time series of joint angles. At the same time, noise that imitates the vibration caused by the elongation of the wires is added for transfer to the actual robot. The results show that the system can be applied to the actual robot RAMIEL as well as to the stable continuous jumping motion in simulation.

Reviewing the previous work of diversity Rein-forcement Learning,diversity is often obtained via an augmented loss function,which requires a balance between reward and diversity.Generally,diversity optimization algorithms use Multi-armed Bandits algorithms to select the coefficient in the pre-defined space. However, the dynamic distribution of reward signals for MABs or the conflict between quality and diversity limits the performance of these methods. We introduce the Phasic Diversity Optimization (PDO) algorithm, a Population-Based Training framework that separates reward and diversity training into distinct phases instead of optimizing a multi-objective function. In the auxiliary phase, agents with poor performance diversified via determinants will not replace the better agents in the archive. The decoupling of reward and diversity allows us to use an aggressive diversity optimization in the auxiliary phase without performance degradation. Furthermore, we construct a dogfight scenario for aerial agents to demonstrate the practicality of the PDO algorithm. We introduce two implementations of PDO archive and conduct tests in the newly proposed adversarial dogfight and MuJoCo simulations. The results show that our proposed algorithm achieves better performance than baselines.

We address in this paper Reinforcement Learning (RL) among agents that are grouped into teams such that there is cooperation within each team but general-sum (non-zero sum) competition across different teams. To develop an RL method that provably achieves a Nash equilibrium, we focus on a linear-quadratic structure. Moreover, to tackle the non-stationarity induced by multi-agent interactions in the finite population setting, we consider the case where the number of agents within each team is infinite, i.e., the mean-field setting. This results in a General-Sum LQ Mean-Field Type Game (GS-MFTGs). We characterize the Nash equilibrium (NE) of the GS-MFTG, under a standard invertibility condition. This MFTG NE is then shown to be $\mathcal{O}(1/M)$-NE for the finite population game where $M$ is a lower bound on the number of agents in each team. These structural results motivate an algorithm called Multi-player Receding-horizon Natural Policy Gradient (MRPG), where each team minimizes its cumulative cost independently in a receding-horizon manner. Despite the non-convexity of the problem, we establish that the resulting algorithm converges to a global NE through a novel problem decomposition into sub-problems using backward recursive discrete-time Hamilton-Jacobi-Isaacs (HJI) equations, in which independent natural policy gradient is shown to exhibit linear convergence under time-independent diagonal dominance. Experiments illuminate the merits of this approach in practice.

This work advances randomized exploration in reinforcement learning (RL) with function approximation modeled by linear mixture MDPs. We establish the first prior-dependent Bayesian regret bound for RL with function approximation; and refine the Bayesian regret analysis for posterior sampling reinforcement learning (PSRL), presenting an upper bound of ${\mathcal{O}}(d\sqrt{H^3 T \log T})$, where $d$ represents the dimensionality of the transition kernel, $H$ the planning horizon, and $T$ the total number of interactions. This signifies a methodological enhancement by optimizing the $\mathcal{O}(\sqrt{\log T})$ factor over the previous benchmark (Osband and Van Roy, 2014) specified to linear mixture MDPs. Our approach, leveraging a value-targeted model learning perspective, introduces a decoupling argument and a variance reduction technique, moving beyond traditional analyses reliant on confidence sets and concentration inequalities to formalize Bayesian regret bounds more effectively.