Trajectory planning under kinodynamic constraints is fundamental for advanced robotics applications that require dexterous, reactive, and rapid skills in complex environments. These constraints, which may represent task, safety, or actuator limitations, are essential for ensuring the proper functioning of robotic platforms and preventing unexpected behaviors. Recent advances in kinodynamic planning demonstrate that learning-to-plan techniques can generate complex and reactive motions under intricate constraints. However, these techniques necessitate the analytical modeling of both the robot and the entire task, a limiting assumption when systems are extremely complex or when constructing accurate task models is prohibitive. This paper addresses this limitation by combining learning-to-plan methods with reinforcement learning, resulting in a novel integration of black-box learning of motion primitives and optimization. We evaluate our approach against state-of-the-art safe reinforcement learning methods, showing that our technique, particularly when exploiting task structure, outperforms baseline methods in challenging scenarios such as planning to hit in robot air hockey. This work demonstrates the potential of our integrated approach to enhance the performance and safety of robots operating under complex kinodynamic constraints.
In today's rapidly evolving military landscape, advancing artificial intelligence (AI) in support of wargaming becomes essential. Despite reinforcement learning (RL) showing promise for developing intelligent agents, conventional RL faces limitations in handling the complexity inherent in combat simulations. This dissertation proposes a comprehensive approach, including targeted observation abstractions, multi-model integration, a hybrid AI framework, and an overarching hierarchical reinforcement learning (HRL) framework. Our localized observation abstraction using piecewise linear spatial decay simplifies the RL problem, enhancing computational efficiency and demonstrating superior efficacy over traditional global observation methods. Our multi-model framework combines various AI methodologies, optimizing performance while still enabling the use of diverse, specialized individual behavior models. Our hybrid AI framework synergizes RL with scripted agents, leveraging RL for high-level decisions and scripted agents for lower-level tasks, enhancing adaptability, reliability, and performance. Our HRL architecture and training framework decomposes complex problems into manageable subproblems, aligning with military decision-making structures. Although initial tests did not show improved performance, insights were gained to improve future iterations. This study underscores AI's potential to revolutionize wargaming, emphasizing the need for continued research in this domain.
Lexicographic multi-objective problems, which impose a lexicographic importance order over the objectives, arise in many real-life scenarios. Existing Reinforcement Learning work directly addressing lexicographic tasks has been scarce. The few proposed approaches were all noted to be heuristics without theoretical guarantees as the Bellman equation is not applicable to them. Additionally, the practical applicability of these prior approaches also suffers from various issues such as not being able to reach the goal state. While some of these issues have been known before, in this work we investigate further shortcomings, and propose fixes for improving practical performance in many cases. We also present a policy optimization approach using our Lexicographic Projection Optimization (LPO) algorithm that has the potential to address these theoretical and practical concerns. Finally, we demonstrate our proposed algorithms on benchmark problems.
In reinforcement learning, conducting task composition by forming cohesive, executable sequences from multiple tasks remains challenging. However, the ability to (de)compose tasks is a linchpin in developing robotic systems capable of learning complex behaviors. Yet, compositional reinforcement learning is beset with difficulties, including the high dimensionality of the problem space, scarcity of rewards, and absence of system robustness after task composition. To surmount these challenges, we view task composition through the prism of category theory -- a mathematical discipline exploring structures and their compositional relationships. The categorical properties of Markov decision processes untangle complex tasks into manageable sub-tasks, allowing for strategical reduction of dimensionality, facilitating more tractable reward structures, and bolstering system robustness. Experimental results support the categorical theory of reinforcement learning by enabling skill reduction, reuse, and recycling when learning complex robotic arm tasks.
Humanoid robots, with their human-like skeletal structure, are especially suited for tasks in human-centric environments. However, this structure is accompanied by additional challenges in locomotion controller design, especially in complex real-world environments. As a result, existing humanoid robots are limited to relatively simple terrains, either with model-based control or model-free reinforcement learning. In this work, we introduce Denoising World Model Learning (DWL), an end-to-end reinforcement learning framework for humanoid locomotion control, which demonstrates the world's first humanoid robot to master real-world challenging terrains such as snowy and inclined land in the wild, up and down stairs, and extremely uneven terrains. All scenarios run the same learned neural network with zero-shot sim-to-real transfer, indicating the superior robustness and generalization capability of the proposed method.
Incorporating inductive biases is a promising approach for tackling challenging robot learning domains with sample-efficient solutions. This paper identifies partially observable domains where symmetries can be a useful inductive bias for efficient learning. Specifically, by encoding the equivariance regarding specific group symmetries into the neural networks, our actor-critic reinforcement learning agents can reuse solutions in the past for related scenarios. Consequently, our equivariant agents outperform non-equivariant approaches significantly in terms of sample efficiency and final performance, demonstrated through experiments on a range of robotic tasks in simulation and real hardware.
Offline multi-agent reinforcement learning (MARL) is increasingly recognized as crucial for effectively deploying RL algorithms in environments where real-time interaction is impractical, risky, or costly. In the offline setting, learning from a static dataset of past interactions allows for the development of robust and safe policies without the need for live data collection, which can be fraught with challenges. Building on this foundational importance, we present EAQ, Episodes Augmentation guided by Q-total loss, a novel approach for offline MARL framework utilizing diffusion models. EAQ integrates the Q-total function directly into the diffusion model as a guidance to maximize the global returns in an episode, eliminating the need for separate training. Our focus primarily lies on cooperative scenarios, where agents are required to act collectively towards achieving a shared goal-essentially, maximizing global returns. Consequently, we demonstrate that our episodes augmentation in a collaborative manner significantly boosts offline MARL algorithm compared to the original dataset, improving the normalized return by +17.3% and +12.9% for medium and poor behavioral policies in SMAC simulator, respectively.
In-context learning (ICL) of large language models (LLMs) has attracted increasing attention in the community where LLMs make predictions only based on instructions augmented with a few examples. Existing example selection methods for ICL utilize sparse or dense retrievers and derive effective performance. However, these methods do not utilize direct feedback of LLM to train the retriever and the examples selected can not necessarily improve the analogy ability of LLM. To tackle this, we propose our policy-based reinforcement learning framework for example selection (RLS), which consists of a language model (LM) selector and an LLM generator. The LM selector encodes the candidate examples into dense representations and selects the top-k examples into the demonstration for LLM. The outputs of LLM are adopted to compute the reward and policy gradient to optimize the LM selector. We conduct experiments on different datasets and significantly outperform existing example selection methods. Moreover, our approach shows advantages over supervised finetuning (SFT) models in few shot setting. Further experiments show the balance of abundance and the similarity with the test case of examples is important for ICL performance of LLM.
Model-based Offline Reinforcement Learning trains policies based on offline datasets and model dynamics, without direct real-world environment interactions. However, this method is inherently challenged by distribution shift. Previous approaches have primarily focused on tackling this issue directly leveraging off-policy mechanisms and heuristic uncertainty in model dynamics, but they resulted in inconsistent objectives and lacked a unified theoretical foundation. This paper offers a comprehensive analysis that disentangles the problem into two key components: model bias and policy shift. We provide both theoretical insights and empirical evidence to demonstrate how these factors lead to inaccuracies in value function estimation and impose implicit restrictions on policy learning. To address these challenges, we derive adjustment terms for model bias and policy shift within a unified probabilistic inference framework. These adjustments are seamlessly integrated into the vanilla reward function to create a novel Shifts-aware Reward (SAR), aiming at refining value learning and facilitating policy training. Furthermore, we introduce Shifts-aware Model-based Offline Reinforcement Learning (SAMBO-RL), a practical framework that efficiently trains classifiers to approximate the SAR for policy optimization. Empirically, we show that SAR effectively mitigates distribution shift, and SAMBO-RL demonstrates superior performance across various benchmarks, underscoring its practical effectiveness and validating our theoretical analysis.
Offline reinforcement learning (RL) learns effective policies from a static target dataset. Despite state-of-the-art (SOTA) offline RL algorithms being promising, they highly rely on the quality of the target dataset. The performance of SOTA algorithms can degrade in scenarios with limited samples in the target dataset, which is often the case in real-world applications. To address this issue, domain adaptation that leverages auxiliary samples from related source datasets (such as simulators) can be beneficial. In this context, determining the optimal way to trade off the source and target datasets remains a critical challenge in offline RL. To the best of our knowledge, this paper proposes the first framework that theoretically and experimentally explores how the weight assigned to each dataset affects the performance of offline RL. We establish the performance bounds and convergence neighborhood of our framework, both of which depend on the selection of the weight. Furthermore, we identify the existence of an optimal weight for balancing the two datasets. All theoretical guarantees and optimal weight depend on the quality of the source dataset and the size of the target dataset. Our empirical results on the well-known Procgen Benchmark substantiate our theoretical contributions.
Data-driven offline reinforcement learning and imitation learning approaches have been gaining popularity in addressing sequential decision-making problems. Yet, these approaches rarely consider learning Pareto-optimal policies from a limited pool of expert datasets. This becomes particularly marked due to practical limitations in obtaining comprehensive datasets for all preferences, where multiple conflicting objectives exist and each expert might hold a unique optimization preference for these objectives. In this paper, we adapt inverse reinforcement learning (IRL) by using reward distance estimates for regularizing the discriminator. This enables progressive generation of a set of policies that accommodate diverse preferences on the multiple objectives, while using only two distinct datasets, each associated with a different expert preference. In doing so, we present a Pareto IRL framework (ParIRL) that establishes a Pareto policy set from these limited datasets. In the framework, the Pareto policy set is then distilled into a single, preference-conditioned diffusion model, thus allowing users to immediately specify which expert's patterns they prefer. Through experiments, we show that ParIRL outperforms other IRL algorithms for various multi-objective control tasks, achieving the dense approximation of the Pareto frontier. We also demonstrate the applicability of ParIRL with autonomous driving in CARLA.
Autonomous vehicles with a self-evolving ability are expected to cope with unknown scenarios in the real-world environment. Take advantage of trial and error mechanism, reinforcement learning is able to self evolve by learning the optimal policy, and it is particularly well suitable for solving decision-making problems. However, reinforcement learning suffers from safety issues and low learning efficiency, especially in the continuous action space. Therefore, the motivation of this paper is to address the above problem by proposing a hybrid Mechanism-Experience-Learning augmented approach. Specifically, to realize the efficient self-evolution, the driving tendency by analogy with human driving experience is proposed to reduce the search space of the autonomous driving problem, while the constrained optimization problem based on a mechanistic model is designed to ensure safety during the self-evolving process. Experimental results show that the proposed method is capable of generating safe and reasonable actions in various complex scenarios, improving the performance of the autonomous driving system. Compared to conventional reinforcement learning, the safety and efficiency of the proposed algorithm are greatly improved. The training process is collision-free, and the training time is equivalent to less than 10 minutes in the real world.
Reinforcement Learning (RL) algorithms suffer from the dependency on accurately engineered reward functions to properly guide the learning agents to do the required tasks. Preference-based reinforcement learning (PbRL) addresses that by utilizing human preferences as feedback from the experts instead of numeric rewards. Due to its promising advantage over traditional RL, PbRL has gained more focus in recent years with many significant advances. In this survey, we present a unified PbRL framework to include the newly emerging approaches that improve the scalability and efficiency of PbRL. In addition, we give a detailed overview of the theoretical guarantees and benchmarking work done in the field, while presenting its recent applications in complex real-world tasks. Lastly, we go over the limitations of the current approaches and the proposed future research directions.
Procedural Content Generation via Reinforcement Learning (PCGRL) has been introduced as a means by which controllable designer agents can be trained based only on a set of computable metrics acting as a proxy for the level's quality and key characteristics. While PCGRL offers a unique set of affordances for game designers, it is constrained by the compute-intensive process of training RL agents, and has so far been limited to generating relatively small levels. To address this issue of scale, we implement several PCGRL environments in Jax so that all aspects of learning and simulation happen in parallel on the GPU, resulting in faster environment simulation; removing the CPU-GPU transfer of information bottleneck during RL training; and ultimately resulting in significantly improved training speed. We replicate several key results from prior works in this new framework, letting models train for much longer than previously studied, and evaluating their behavior after 1 billion timesteps. Aiming for greater control for human designers, we introduce randomized level sizes and frozen "pinpoints" of pivotal game tiles as further ways of countering overfitting. To test the generalization ability of learned generators, we evaluate models on large, out-of-distribution map sizes, and find that partial observation sizes learn more robust design strategies.
Reinforcement learning techniques leveraging deep learning have made tremendous progress in recent years. However, the complexity of neural networks prevents practitioners from understanding their behavior. Decision trees have gained increased attention in supervised learning for their inherent interpretability, enabling modelers to understand the exact prediction process after learning. This paper considers the problem of optimizing interpretable decision tree policies to replace neural networks in reinforcement learning settings. Previous works have relaxed the tree structure, restricted to optimizing only tree leaves, or applied imitation learning techniques to approximately copy the behavior of a neural network policy with a decision tree. We propose the Decision Tree Policy Optimization (DTPO) algorithm that directly optimizes the complete decision tree using policy gradients. Our technique uses established decision tree heuristics for regression to perform policy optimization. We empirically show that DTPO is a competitive algorithm compared to imitation learning algorithms for optimizing decision tree policies in reinforcement learning.
Utilizing deep learning models to learn part-based representations holds significant potential for interpretable-by-design approaches, as these models incorporate latent causes obtained from feature representations through simple addition. However, training a part-based learning model presents challenges, particularly in enforcing non-negative constraints on the model's parameters, which can result in training difficulties such as instability and convergence issues. Moreover, applying such approaches in Deep Reinforcement Learning (RL) is even more demanding due to the inherent instabilities that impact many optimization methods. In this paper, we propose a non-negative training approach for actor models in RL, enabling the extraction of part-based representations that enhance interpretability while adhering to non-negative constraints. To this end, we employ a non-negative initialization technique, as well as a modified sign-preserving training method, which can ensure better gradient flow compared to existing approaches. We demonstrate the effectiveness of the proposed approach using the well-known Cartpole benchmark.
In the face of difficult exploration problems in reinforcement learning, we study whether giving an agent an object-centric mapping (describing a set of items and their attributes) allow for more efficient learning. We found this problem is best solved hierarchically by modelling items at a higher level of state abstraction to pixels, and attribute change at a higher level of temporal abstraction to primitive actions. This abstraction simplifies the transition dynamic by making specific future states easier to predict. We make use of this to propose a fully model-based algorithm that learns a discriminative world model, plans to explore efficiently with only a count-based intrinsic reward, and can subsequently plan to reach any discovered (abstract) states. We demonstrate the model's ability to (i) efficiently solve single tasks, (ii) transfer zero-shot and few-shot across item types and environments, and (iii) plan across long horizons. Across a suite of 2D crafting and MiniHack environments, we empirically show our model significantly out-performs state-of-the-art low-level methods (without abstraction), as well as performant model-free and model-based methods using the same abstraction. Finally, we show how to learn low level object-perturbing policies via reinforcement learning, and the object mapping itself by supervised learning.
As urban residents demand higher travel quality, vehicle dispatch has become a critical component of online ride-hailing services. However, current vehicle dispatch systems struggle to navigate the complexities of urban traffic dynamics, including unpredictable traffic conditions, diverse driver behaviors, and fluctuating supply and demand patterns. These challenges have resulted in travel difficulties for passengers in certain areas, while many drivers in other areas are unable to secure orders, leading to a decline in the overall quality of urban transportation services. To address these issues, this paper introduces GARLIC: a framework of GPT-Augmented Reinforcement Learning with Intelligent Control for vehicle dispatching. GARLIC utilizes multiview graphs to capture hierarchical traffic states, and learns a dynamic reward function that accounts for individual driving behaviors. The framework further integrates a GPT model trained with a custom loss function to enable high-precision predictions and optimize dispatching policies in real-world scenarios. Experiments conducted on two real-world datasets demonstrate that GARLIC effectively aligns with driver behaviors while reducing the empty load rate of vehicles.
The advancement of Offline Reinforcement Learning (RL) and Offline Multi-Agent Reinforcement Learning (MARL) critically depends on the availability of high-quality, pre-collected offline datasets that represent real-world complexities and practical applications. However, existing datasets often fall short in their simplicity and lack of realism. To address this gap, we propose Hokoff, a comprehensive set of pre-collected datasets that covers both offline RL and offline MARL, accompanied by a robust framework, to facilitate further research. This data is derived from Honor of Kings, a recognized Multiplayer Online Battle Arena (MOBA) game known for its intricate nature, closely resembling real-life situations. Utilizing this framework, we benchmark a variety of offline RL and offline MARL algorithms. We also introduce a novel baseline algorithm tailored for the inherent hierarchical action space of the game. We reveal the incompetency of current offline RL approaches in handling task complexity, generalization and multi-task learning.
In this paper, we study reinforcement learning in Markov Decision Processes with Probabilistic Reward Machines (PRMs), a form of non-Markovian reward commonly found in robotics tasks. We design an algorithm for PRMs that achieves a regret bound of $\widetilde{O}(\sqrt{HOAT} + H^2O^2A^{3/2} + H\sqrt{T})$, where $H$ is the time horizon, $O$ is the number of observations, $A$ is the number of actions, and $T$ is the number of time-steps. This result improves over the best-known bound, $\widetilde{O}(H\sqrt{OAT})$ of \citet{pmlr-v206-bourel23a} for MDPs with Deterministic Reward Machines (DRMs), a special case of PRMs. When $T \geq H^3O^3A^2$ and $OA \geq H$, our regret bound leads to a regret of $\widetilde{O}(\sqrt{HOAT})$, which matches the established lower bound of $\Omega(\sqrt{HOAT})$ for MDPs with DRMs up to a logarithmic factor. To the best of our knowledge, this is the first efficient algorithm for PRMs. Additionally, we present a new simulation lemma for non-Markovian rewards, which enables reward-free exploration for any non-Markovian reward given access to an approximate planner. Complementing our theoretical findings, we show through extensive experiment evaluations that our algorithm indeed outperforms prior methods in various PRM environments.