Recently, graph-based planning algorithms have gained much attention to solve goal-conditioned reinforcement learning (RL) tasks: they provide a sequence of subgoals to reach the target-goal, and the agents learn to execute subgoal-conditioned policies. However, the sample-efficiency of such RL schemes still remains a challenge, particularly for long-horizon tasks. To address this issue, we present a simple yet effective self-imitation scheme which distills a subgoal-conditioned policy into the target-goal-conditioned policy. Our intuition here is that to reach a target-goal, an agent should pass through a subgoal, so target-goal- and subgoal- conditioned policies should be similar to each other. We also propose a novel scheme of stochastically skipping executed subgoals in a planned path, which further improves performance. Unlike prior methods that only utilize graph-based planning in an execution phase, our method transfers knowledge from a planner along with a graph into policy learning. We empirically show that our method can significantly boost the sample-efficiency of the existing goal-conditioned RL methods under various long-horizon control tasks.

Recent rapid developments in reinforcement learning algorithms have been giving us novel possibilities in many fields. However, due to their exploring property, we have to take the risk into consideration when we apply those algorithms to safety-critical problems especially in real environments. In this study, we deal with a safe exploration problem in reinforcement learning under the existence of disturbance. We define the safety during learning as satisfaction of the constraint conditions explicitly defined in terms of the state and propose a safe exploration method that uses partial prior knowledge of a controlled object and disturbance. The proposed method assures the satisfaction of the explicit state constraints with a pre-specified probability even if the controlled object is exposed to a stochastic disturbance following a normal distribution. As theoretical results, we introduce sufficient conditions to construct conservative inputs not containing an exploring aspect used in the proposed method and prove that the safety in the above explained sense is guaranteed with the proposed method. Furthermore, we illustrate the validity and effectiveness of the proposed method through numerical simulations of an inverted pendulum and a four-bar parallel link robot manipulator.

Object rearrangement is a challenge for embodied agents because solving these tasks requires generalizing across a combinatorially large set of configurations of entities and their locations. Worse, the representations of these entities are unknown and must be inferred from sensory percepts. We present a hierarchical abstraction approach to uncover these underlying entities and achieve combinatorial generalization from unstructured visual inputs. By constructing a factorized transition graph over clusters of entity representations inferred from pixels, we show how to learn a correspondence between intervening on states of entities in the agent's model and acting on objects in the environment. We use this correspondence to develop a method for control that generalizes to different numbers and configurations of objects, which outperforms current offline deep RL methods when evaluated on simulated rearrangement tasks.

We study the offline contextual bandit problem, where we aim to acquire an optimal policy using observational data. However, this data usually contains two deficiencies: (i) some variables that confound actions are not observed, and (ii) missing observations exist in the collected data. Unobserved confounders lead to a confounding bias and missing observations cause bias and inefficiency problems. To overcome these challenges and learn the optimal policy from the observed dataset, we present a new algorithm called Causal-Adjusted Pessimistic (CAP) policy learning, which forms the reward function as the solution of an integral equation system, builds a confidence set, and greedily takes action with pessimism. With mild assumptions on the data, we develop an upper bound to the suboptimality of CAP for the offline contextual bandit problem.

In reward-free reinforcement learning (RL), an agent explores the environment first without any reward information, in order to achieve certain learning goals afterwards for any given reward. In this paper we focus on reward-free RL under low-rank MDP models, in which both the representation and linear weight vectors are unknown. Although various algorithms have been proposed for reward-free low-rank MDPs, the corresponding sample complexity is still far from being satisfactory. In this work, we first provide the first known sample complexity lower bound that holds for any algorithm under low-rank MDPs. This lower bound implies it is strictly harder to find a near-optimal policy under low-rank MDPs than under linear MDPs. We then propose a novel model-based algorithm, coined RAFFLE, and show it can both find an $\epsilon$-optimal policy and achieve an $\epsilon$-accurate system identification via reward-free exploration, with a sample complexity significantly improving the previous results. Such a sample complexity matches our lower bound in the dependence on $\epsilon$, as well as on $K$ in the large $d$ regime, where $d$ and $K$ respectively denote the representation dimension and action space cardinality. Finally, we provide a planning algorithm (without further interaction with true environment) for RAFFLE to learn a near-accurate representation, which is the first known representation learning guarantee under the same setting.

Evaluating the worst-case performance of a reinforcement learning (RL) agent under the strongest/optimal adversarial perturbations on state observations (within some constraints) is crucial for understanding the robustness of RL agents. However, finding the optimal adversary is challenging, in terms of both whether we can find the optimal attack and how efficiently we can find it. Existing works on adversarial RL either use heuristics-based methods that may not find the strongest adversary, or directly train an RL-based adversary by treating the agent as a part of the environment, which can find the optimal adversary but may become intractable in a large state space. This paper introduces a novel attacking method to find the optimal attacks through collaboration between a designed function named "actor" and an RL-based learner named "director". The actor crafts state perturbations for a given policy perturbation direction, and the director learns to propose the best policy perturbation directions. Our proposed algorithm, PA-AD, is theoretically optimal and significantly more efficient than prior RL-based works in environments with large state spaces. Empirical results show that our proposed PA-AD universally outperforms state-of-the-art attacking methods in various Atari and MuJoCo environments. By applying PA-AD to adversarial training, we achieve state-of-the-art empirical robustness in multiple tasks under strong adversaries. The codebase is released at https://github.com/umd-huang-lab/paad_adv_rl.

Reinforcement learning (RL) has achieved enormous progress in solving various sequential decision-making problems, such as control tasks in robotics. Since policies are overfitted to training environments, RL methods have often failed to be generalized to safety-critical test scenarios. Robust adversarial RL (RARL) was previously proposed to train an adversarial network that applies disturbances to a system, which improves the robustness in test scenarios. However, an issue of neural network-based adversaries is that integrating system requirements without handcrafting sophisticated reward signals are difficult. Safety falsification methods allow one to find a set of initial conditions and an input sequence, such that the system violates a given property formulated in temporal logic. In this paper, we propose falsification-based RARL (FRARL): this is the first generic framework for integrating temporal logic falsification in adversarial learning to improve policy robustness. By applying our falsification method, we do not need to construct an extra reward function for the adversary. Moreover, we evaluate our approach on a braking assistance system and an adaptive cruise control system of autonomous vehicles. Our experimental results demonstrate that policies trained with a falsification-based adversary generalize better and show less violation of the safety specification in test scenarios than those trained without an adversary or with an adversarial network.

The olfactory search POMDP (partially observable Markov decision process) is a sequential decision-making problem designed to mimic the task faced by insects searching for a source of odor in turbulence, and its solutions have applications to sniffer robots. As exact solutions are out of reach, the challenge consists in finding the best possible approximate solutions while keeping the computational cost reasonable. We provide a quantitative benchmarking of a solver based on deep reinforcement learning against traditional POMDP approximate solvers. We show that deep reinforcement learning is a competitive alternative to standard methods, in particular to generate lightweight policies suitable for robots.

One of the essential missions in the AI research community is to build an autonomous embodied agent that can attain high-level performance across a wide spectrum of tasks. However, acquiring reward/penalty in all open-ended tasks is unrealistic, making the Reinforcement Learning (RL) training procedure impossible. In this paper, we propose a novel cross-modal contrastive learning framework architecture, CLIP4MC, aiming to learn an RL-friendly vision-language model that serves as a reward function for open-ended tasks. Therefore, no further task-specific reward design is needed. Intuitively, it is more reasonable for the model to address the similarity between the video snippet and the language prompt at both the action and entity levels. To this end, a motion encoder is proposed to capture the motion embeddings across different intervals. The correlation scores are then used to construct the auxiliary reward signal for RL agents. Moreover, we construct a neat YouTube dataset based on the large-scale YouTube database provided by MineDojo. Specifically, two rounds of filtering operations guarantee that the dataset covers enough essential information and that the video-text pair is highly correlated. Empirically, we show that the proposed method achieves better performance on RL tasks compared with baselines.

Building systems that are good for society in the face of complex societal effects requires a dynamic approach. Recent approaches to machine learning (ML) documentation have demonstrated the promise of discursive frameworks for deliberation about these complexities. However, these developments have been grounded in a static ML paradigm, leaving the role of feedback and post-deployment performance unexamined. Meanwhile, recent work in reinforcement learning has shown that the effects of feedback and optimization objectives on system behavior can be wide-ranging and unpredictable. In this paper we sketch a framework for documenting deployed and iteratively updated learning systems, which we call Reward Reports. Taking inspiration from various contributions to the technical literature on reinforcement learning, we outline Reward Reports as living documents that track updates to design choices and assumptions behind what a particular automated system is optimizing for. They are intended to track dynamic phenomena arising from system deployment, rather than merely static properties of models or data. After presenting the elements of a Reward Report, we discuss a concrete example: Meta's BlenderBot 3 chatbot. Several others for game-playing (DeepMind's MuZero), content recommendation (MovieLens), and traffic control (Project Flow) are included in the appendix.