Abstract--Curriculum learning is gaining popularity in (deep) reinforcement learning. It can alleviate the burden on data collection and provide better exploration policies through transfer and generalization from less complex tasks. Current methods for automatic task sequencing for curriculum learning in reinforcement learning provided initial heuristic solutions, with little to no guarantee on their quality. We introduce an optimization framework for task sequencing composed of the problem definition, several candidate performance metrics for optimization, and three benchmark algorithms. We experimentally show that the two most commonly used baselines (learning with no curriculum, and with a random curriculum) perform worse than a simple greedy algorithm. Furthermore, we show theoretically and demonstrate experimentally that the three proposed algorithms provide increasing solution quality at moderately increasing computational complexity, and show that they constitute better baselines for curriculum learning in reinforcement learning. Reinforcement Learning (RL) has recently been successfully applied to a number of tasks whose complexity would have appeared overwhelming only a few years ago [1], [2]. In such large and complex environments, classical exploration strategies designed for Markov Decision Processes (MDPs), aiming at visiting every state the most efficiently, are inadequate. One approach actively investigated is the use of transfer learning [3] to generalize from previous similar tasks, and more recently the application of transfer learning to sequences of tasks of increasing complexity forming a curriculum . Curriculum Learning is often employed in (Deep) RL [4], [5] to let the agent progress more quickly towards better behaviors, but curricula are mostly designed by hand. Curriculum learning has the potential to greatly increase the quality of the behavior discovered by the agent. However, at the moment, creating an appropriate curriculum requires significant human intuition.

Many recent successful (deep) reinforcement learning algorithms make use of regularization, generally based on entropy or on Kullback-Leibler divergence. We propose a general theory of regularized Markov Decision Processes that generalizes these approaches in two directions: we consider a larger class of regularizers, and we consider the general modified policy iteration approach, encompassing both policy iteration and value iteration. The core building blocks of this theory are a notion of regularized Bellman operator and the Legendre-Fenchel transform, a classical tool of convex optimization. This approach allows for error propagation analyses of general algorithmic schemes of which (possibly variants of) classical algorithms such as Trust Region Policy Optimization, Soft Q-learning, Stochastic Actor Critic or Dynamic Policy Programming are special cases. This also draws connections to proximal convex optimization, especially to Mirror Descent.

In this paper, we propose an unsupervised reinforcement learning agent called CLIC for Curriculum Learning and Imitation for Control. This agent learns to control features in its environment without external rewards, and observes the actions of a third party agent, Bob, who does not necessarily provide explicit guidance. CLIC selects which feature to train on and what to imitate from Bob's behavior by maximizing its learning progress. We show that CLIC can effectively identify helpful behaviors in Bob's actions, and imitate them to control the environment faster. CLIC can also follow Bob when he acts as a mentor and provides ordered demonstrations. Finally, when Bob controls features than the agent cannot, or in presence of a hierarchy between aspects of the environment, we show that CLIC ignores non-reproducible and already mastered behaviors, resulting in a greater benefit from imitation.

We present a novel method for learning a set of disentangled reward functions that sum to the original environment reward and are constrained to be independently obtainable. We define independent obtainability in terms of value functions with respect to obtaining one learned reward while pursuing another learned reward. Empirically, we illustrate that our method can learn meaningful reward decompositions in a variety of domains and that these decompositions exhibit some form of generalization performance when the environment's reward is modified. Theoretically, we derive results about the effect of maximizing our method's objective on the resulting reward functions and their corresponding optimal policies.

Reinforcement Learning algorithms typically require millions of environment interactions to learn successful policies in sparse reward settings. Hindsight Experience Replay (HER) was introduced as a technique to increase sample efficiency through re-imagining unsuccessful trajectories as successful ones by replacing the originally intended goals. However, this method is not applicable to visual domains where the goal configuration is unknown and must be inferred from observation. In this work, we show how unsuccessful visual trajectories can be hallucinated to be successful using a generative model trained on relatively few snapshots of the goal. As far as we are aware, this is the first work that does so with the agent policy conditioned solely on its state. We then apply this model to training reinforcement learning agents in discrete and continuous settings. We show results on a navigation and pick-and-place task in a 3D environment and on a simulated robotics application. Our method shows marked improvement over standard RL algorithms and baselines derived from prior work.

A grand challenge in reinforcement learning is intelligent exploration, especially when rewards are sparse or deceptive. Two Atari games serve as benchmarks for such hard-exploration domains: Montezuma's Revenge and Pitfall. On both games, current RL algorithms perform poorly, even those with intrinsic motivation, which is the dominant method to improve performance on hard-exploration domains. To address this shortfall, we introduce a new algorithm called Go-Explore. It exploits the following principles: (1) remember previously visited states, (2) first return to a promising state (without exploration), then explore from it, and (3) solve simulated environments through any available means (including by introducing determinism), then robustify via imitation learning. The combined effect of these principles is a dramatic performance improvement on hard-exploration problems. On Montezuma's Revenge, Go-Explore scores a mean of over 43k points, almost 4 times the previous state of the art. Go-Explore can also harness human-provided domain knowledge and, when augmented with it, scores a mean of over 650k points on Montezuma's Revenge. Its max performance of nearly 18 million surpasses the human world record, meeting even the strictest definition of "superhuman" performance. On Pitfall, Go-Explore with domain knowledge is the first algorithm to score above zero. Its mean score of almost 60k points exceeds expert human performance. Because Go-Explore produces high-performing demonstrations automatically and cheaply, it also outperforms imitation learning work where humans provide solution demonstrations. Go-Explore opens up many new research directions into improving it and weaving its insights into current RL algorithms. It may also enable progress on previously unsolvable hard-exploration problems in many domains, especially those that harness a simulator during training (e.g. robotics).

The goal of this paper is to provide a unifying view of a wide range of problems of interest in machine learning by framing them as the minimization of functionals defined on the space of probability measures. In particular, we show that generative adversarial networks, variational inference, and actor-critic methods in reinforcement learning can all be seen through the lens of our framework. We then discuss a generic optimization algorithm for our formulation, called probability functional descent (PFD), and show how this algorithm recovers existing methods developed independently in the settings mentioned earlier.

The ability to transfer skills across tasks has the potential to scale up reinforcement learning (RL) agents to environments currently out of reach. Recently, a framework based on two ideas, successor features (SFs) and generalised policy improvement (GPI), has been introduced as a principled way of transferring skills. In this paper we extend the SFs & GPI framework in two ways. One of the basic assumptions underlying the original formulation of SFs & GPI is that rewards for all tasks of interest can be computed as linear combinations of a fixed set of features. We relax this constraint and show that the theoretical guarantees supporting the framework can be extended to any set of tasks that only differ in the reward function. Our second contribution is to show that one can use the reward functions themselves as features for future tasks, without any loss of expressiveness, thus removing the need to specify a set of features beforehand. This makes it possible to combine SFs & GPI with deep learning in a more stable way. We empirically verify this claim on a complex 3D environment where observations are images from a first-person perspective. We show that the transfer promoted by SFs & GPI leads to very good policies on unseen tasks almost instantaneously. We also describe how to learn policies specialised to the new tasks in a way that allows them to be added to the agent's set of skills, and thus be reused in the future.

Navigation research is attracting renewed interest with the advent of learning-based methods. However, this new line of work is largely disconnected from well-established classic navigation approaches. In this paper, we take a step towards coordinating these two directions of research. We set up classic and learning-based navigation systems in common simulated environments and thoroughly evaluate them in indoor spaces of varying complexity, with access to different sensory modalities. Additionally, we measure human performance in the same environments. We find that a classic pipeline, when properly tuned, can perform very well in complex cluttered environments. On the other hand, learned systems can operate more robustly with a limited sensor suite. Both approaches are still far from human-level performance.

The Researchers from the University of California, Berkeley, have launched a start-up, Embodied Intelligence, Inc., to use the latest techniques of deep reinforcement learning and artificial intelligence to make industrial robots easily teachable. Robots today must be programmed by writing computer code, but imagine donning a VR headset and virtually guiding a robot through a task, like you would move the arms of a puppet, and then letting the robot take it from there. That's the vision of Pieter Abbeel, a professor of electrical engineering and computer science at the University of California, Berkeley, and his students, Peter Chen, Rocky Duan and Tianhao Zhang, who have launched a startup, Embodied Intelligence Inc., to use the latest techniques of deep reinforcement learning and artificial intelligence to make industrial robots easily teachable. "Right now, if you want to set up a robot, you program that robot to do what you want it to do, which takes a lot of time and a lot of expertise," said Abbeel, who is currently on leave to turn his vision into reality. "With our advances in machine learning, we can write a piece of software once -- machine learning code that enables the robot to learn -- and then when the robot needs to be equipped with a new skill, we simply provide new data."