The team's algorithm, called Dreamer, uses past experiences to build up a model of the surrounding world. Dreamer also allows the robot to conduct trial-and-error calculations in a computer program as opposed to the real world, by predicting potential future outcomes of its potential actions. This allows it to learn faster than it could purely by doing. Once the robot had learned to walk, it kept learning to adapt to unexpected situations, such as resisting being toppled by a stick. "Teaching robots through trial and error is a difficult problem, made even harder by the long training times such teaching requires," says Lerrel Pinto, an assistant professor of computer science at New York University, who specializes in robotics and machine learning.

We study the problem of online learning in adversarial bandit problems under a partial observability model called off-policy feedback. In this sequential decision making problem, the learner cannot directly observe its rewards, but instead sees the ones obtained by another unknown policy run in parallel (behavior policy). Instead of a standard exploration-exploitation dilemma, the learner has to face another challenge in this setting: due to limited observations outside of their control, the learner may not be able to estimate the value of each policy equally well. To address this issue, we propose a set of algorithms that guarantee regret bounds that scale with a natural notion of mismatch between any comparator policy and the behavior policy, achieving improved performance against comparators that are well-covered by the observations. We also provide an extension to the setting of adversarial linear contextual bandits, and verify the theoretical guarantees via a set of experiments. Our key algorithmic idea is adapting the notion of pessimistic reward estimators that has been recently popular in the context of off-policy reinforcement learning.

We introduce a new distributed policy gradient algorithm and show that it outperforms existing reward-aware training procedures such as REINFORCE, minimum risk training (MRT) and proximal policy optimization (PPO) in terms of training stability and generalization performance when optimizing machine translation models. Our algorithm, which we call MAD (on account of using the mean absolute deviation in the importance weighting calculation), has distributed data generators sampling multiple candidates per source sentence on worker nodes, while a central learner updates the policy. MAD depends crucially on two variance reduction strategies: (1) a conditional reward normalization method that ensures each source sentence has both positive and negative reward translation examples and (2) a new robust importance weighting scheme that acts as a conditional entropy regularizer. Experiments on a variety of translation tasks show that policies learned using the MAD algorithm perform very well when using both greedy decoding and beam search, and that the learned policies are sensitive to the specific reward used during training.

Abstract-- Robots often face situations where grasping a goal object is desirable but not feasible due to other present objects preventing the grasp action. We present a deep Reinforcement Learning approach to learn grasping and pushing policies for manipulating a goal object in highly cluttered environments to address this problem. In particular, a dual Reinforcement Learning model approach is proposed, which presents high resilience in handling complicated scenes, reaching an average of 98% task completion using primitive objects in a simulation environment. To evaluate the performance of the proposed approach, we performed two extensive sets of experiments in packed objects and a pile of object scenarios with a total of 1000 test runs in simulation. Experimental results showed that the proposed method worked very well in both scenarios and outperformed the recent state-of-the-art approaches. Demo video, trained models, and source code for the results reproducibility purpose are publicly available. Grasping is a crucial skill for many robotics applications. It further opens opportunities to robots to numerous skills that are not This paper mainly focuses on self-supervised learning for attainable without a versed object's grasp. In many scenarios, pushing and grasping objects in highly cluttered environments objects do not appear in isolation, and therefore, grasping via deep Reinforcement Learning (RL).

The improvement of traffic efficiency at urban intersections receives strong research interest in the field of automated intersection management. So far, mostly non-learning algorithms like reservation or optimization-based ones were proposed to solve the underlying multi-agent planning problem. At the same time, automated driving functions for a single ego vehicle are increasingly implemented using machine learning methods. In this work, we build upon a previously presented graph-based scene representation and graph neural network to approach the problem using reinforcement learning. The scene representation is improved in key aspects by using edge features in addition to the existing node features for the vehicles. This leads to an increased representation quality that is leveraged by an updated network architecture. The paper provides an in-depth evaluation of the proposed method against baselines that are commonly used in automatic intersection management. Compared to a traditional signalized intersection and an enhanced first-in-first-out scheme, a significant reduction of induced delay is observed at varying traffic densities. Finally, the generalization capability of the graph-based representation is evaluated by testing the policy on intersection layouts not seen during training. The model generalizes virtually without restrictions to smaller intersection layouts and within certain limits to larger ones.

Learning from previously collected datasets of expert data offers the promise of acquiring robotic policies without unsafe and costly online explorations. However, a major challenge is a distributional shift between the states in the training dataset and the ones visited by the learned policy at the test time. While prior works mainly studied the distribution shift caused by the policy during the offline training, the problem of recovering from out-of-distribution states at the deployment time is not very well studied yet. We alleviate the distributional shift at the deployment time by introducing a recovery policy that brings the agent back to the training manifold whenever it steps out of the in-distribution states, e.g., due to an external perturbation. The recovery policy relies on an approximation of the training data density and a learned equivariant mapping that maps visual observations into a latent space in which translations correspond to the robot actions. We demonstrate the effectiveness of the proposed method through several manipulation experiments on a real robotic platform. Our results show that the recovery policy enables the agent to complete tasks while the behavioral cloning alone fails because of the distributional shift problem.

Deep reinforcement learning (DRL) has attracted much attention in automated game testing. Early attempts rely on game internal information for game space exploration, thus requiring deep integration with games, which is inconvenient for practical applications. In this work, we propose using only screenshots/pixels as input for automated game testing and build a general game testing agent, Inspector, that can be easily applied to different games without deep integration with games. In addition to covering all game space for testing, our agent tries to take human-like behaviors to interact with key objects in a game, since some bugs usually happen in player-object interactions. Inspector is based on purely pixel inputs and comprises three key modules: game space explorer, key object detector, and human-like object investigator. Game space explorer aims to explore the whole game space by using a curiosity-based reward function with pixel inputs. Key object detector aims to detect key objects in a game, based on a small number of labeled screenshots. Human-like object investigator aims to mimic human behaviors for investigating key objects via imitation learning. We conduct experiments on two popular video games: Shooter Game and Action RPG Game. Experiment results demonstrate the effectiveness of Inspector in exploring game space, detecting key objects, and investigating objects. Moreover, Inspector successfully discovers two potential bugs in those two games. The demo video of Inspector is available at https://github.com/Inspector-GameTesting/Inspector-GameTesting.

Explainability of Reinforcement Learning (RL) policies remains a challenging research problem, particularly when considering RL in a safety context. Understanding the decisions and intentions of an RL policy offer avenues to incorporate safety into the policy by limiting undesirable actions. We propose the use of a Boolean Decision Rules model to create a post-hoc rule-based summary of an agent's policy. We evaluate our proposed approach using a DQN agent trained on an implementation of a lava gridworld and show that, given a hand-crafted feature representation of this gridworld, simple generalised rules can be created, giving a post-hoc explainable summary of the agent's policy. We discuss possible avenues to introduce safety into a RL agent's policy by using rules generated by this rule-based model as constraints imposed on the agent's policy, as well as discuss how creating simple rule summaries of an agent's policy may help in the debugging process of RL agents.

The action governor is an add-on scheme to a nominal control loop that monitors and adjusts the control actions to enforce safety specifications expressed as pointwise-in-time state and control constraints. In this paper, we introduce the Robust Action Governor (RAG) for systems the dynamics of which can be represented using discrete-time Piecewise Affine (PWA) models with both parametric and additive uncertainties and subject to non-convex constraints. We develop the theoretical properties and computational approaches for the RAG. After that, we introduce the use of the RAG for realizing safe Reinforcement Learning (RL), i.e., ensuring all-time constraint satisfaction during online RL exploration-and-exploitation process. This development enables safe real-time evolution of the control policy and adaptation to changes in the operating environment and system parameters (due to aging, damage, etc.). We illustrate the effectiveness of the RAG in constraint enforcement and safe RL using the RAG by considering their applications to a soft-landing problem of a mass-spring-damper system.

In contrast to a standard closed-set domain adaptation task, partial domain adaptation setup caters to a realistic scenario by relaxing the identical label set assumption. The fact of source label set subsuming the target label set, however, introduces few additional obstacles as training on private source category samples thwart relevant knowledge transfer and mislead the classification process. To mitigate these issues, we devise a mechanism for strategic selection of highly-confident target samples essential for the estimation of class-importance weights. Furthermore, we capture class-discriminative and domain-invariant features by coupling the process of achieving compact and distinct class distributions with an adversarial objective. Experimental findings over numerous cross-domain classification tasks demonstrate the potential of the proposed technique to deliver superior and comparable accuracy over existing methods.