In this paper, we introduce a novel online model-based reinforcement learning algorithm that uses Unscented Transform to propagate uncertainty for the prediction of the future reward. Previous approaches either approximate the state distribution at each step of the prediction horizon with a Gaussian, or perform Monte Carlo simulations to estimate the rewards. Our method, depending on the number of sigma points employed, can propagate either mean and covariance with minimal points, or higher-order moments with more points similarly to Monte Carlo. The whole framework is implemented as a computational graph for online training. Furthermore, in order to prevent explosion in the number of sigma points when propagating through a generic state-dependent uncertainty model, we add sigma-point expansion and contraction layers to our graph, which are designed using the principle of moment matching. Finally, we propose gradient descent inspired by Sequential Quadratic Programming to update policy parameters in the presence of state constraints. We demonstrate the proposed method with two applications in simulation. The first one designs a stabilizing controller for the cart-pole problem when the dynamics is known with state-dependent uncertainty. The second example, following up on our previous work, tunes the parameters of a control barrier function-based Quadratic Programming controller for a leader-follower problem in the presence of input constraints.

The endeavor of artificial intelligence (AI) is to design autonomous agents capable of achieving complex tasks. Namely, reinforcement learning (RL) proposes a theoretical background to learn optimal behaviors. In practice, RL algorithms rely on geometric discounts to evaluate this optimality. Unfortunately, this does not cover decision processes where future returns are not exponentially less valuable. Depending on the problem, this limitation induces sample-inefficiency (as feed-backs are exponentially decayed) and requires additional curricula/exploration mechanisms (to deal with sparse, deceptive or adversarial rewards). In this paper, we tackle these issues by generalizing the discounted problem formulation with a family of delayed objective functions. We investigate the underlying RL problem to derive: 1) the optimal stationary solution and 2) an approximation of the optimal non-stationary control. The devised algorithms solved hard exploration problems on tabular environment and improved sample-efficiency on classic simulated robotics benchmarks.

Learning to manipulate 3D objects in an interactive environment has been a challenging problem in Reinforcement Learning (RL). In particular, it is hard to train a policy that can generalize over objects with different semantic categories, diverse shape geometry and versatile functionality. Recently, the technique of visual affordance has shown great prospects in providing object-centric information priors with effective actionable semantics. As such, an effective policy can be trained to open a door by knowing how to exert force on the handle. However, to learn the affordance, it often requires human-defined action primitives, which limits the range of applicable tasks. In this study, we take advantage of visual affordance by using the contact information generated during the RL training process to predict contact maps of interest. Such contact prediction process then leads to an end-to-end affordance learning framework that can generalize over different types of manipulation tasks. Surprisingly, the effectiveness of such framework holds even under the multi-stage and the multi-agent scenarios. We tested our method on eight types of manipulation tasks. Results showed that our methods outperform baseline algorithms, including visual-based affordance methods and RL methods, by a large margin on the success rate. The demonstration can be found at https://sites.google.com/view/rlafford/.

Offline Reinforcement Learning has attracted much interest in solving the application challenge for traditional reinforcement learning. Offline reinforcement learning uses previously-collected datasets to train agents without any interaction. For addressing the overestimation of OOD (out-of-distribution) actions, conservative estimates give a low value for all inputs. Previous conservative estimation methods are usually difficult to avoid the impact of OOD actions on Q-value estimates. In addition, these algorithms usually need to lose some computational efficiency to achieve the purpose of conservative estimation. In this paper, we propose a simple conservative estimation method, double conservative estimates (DCE), which use two conservative estimation method to constraint policy. Our algorithm introduces V-function to avoid the error of in-distribution action while implicit achieving conservative estimation. In addition, our algorithm uses a controllable penalty term changing the degree of conservatism in training. We theoretically show how this method influences the estimation of OOD actions and in-distribution actions. Our experiment separately shows that two conservative estimation methods impact the estimation of all state-action. DCE demonstrates the state-of-the-art performance on D4RL.

Reinforcement learning algorithms have become more complex since the invention of target networks. Unfortunately, target networks have not kept up with this increased complexity, instead requiring approximate solutions to be computationally feasible. These approximations increase noise in the Q-value targets and in the replay sampling distribution. Paused Agent Replay Refresh (PARR) is a drop-in replacement for target networks that supports more complex learning algorithms without this need for approximation. Using a basic Q-network architecture, and refreshing the novelty values, target values, and replay sampling distribution, PARR gets 2500 points in Montezuma's Revenge after only 30.9 million Atari frames. Finally, interpreting PARR in the context of carbon-based learning offers a new reason for sleep.

The common approach for local navigation on challenging environments with legged robots requires path planning, path following and locomotion, which usually requires a locomotion control policy that accurately tracks a commanded velocity. However, by breaking down the navigation problem into these sub-tasks, we limit the robot's capabilities since the individual tasks do not consider the full solution space. In this work, we propose to solve the complete problem by training an end-to-end policy with deep reinforcement learning. Instead of continuously tracking a precomputed path, the robot needs to reach a target position within a provided time. The task's success is only evaluated at the end of an episode, meaning that the policy does not need to reach the target as fast as possible. It is free to select its path and the locomotion gait. Training a policy in this way opens up a larger set of possible solutions, which allows the robot to learn more complex behaviors. We compare our approach to velocity tracking and additionally show that the time dependence of the task reward is critical to successfully learn these new behaviors. Finally, we demonstrate the successful deployment of policies on a real quadrupedal robot. The robot is able to cross challenging terrains, which were not possible previously, while using a more energy-efficient gait and achieving a higher success rate.

Successful Artificial Intelligence systems often require numerous labeled data to extract information from document images. In this paper, we investigate the problem of improving the performance of Artificial Intelligence systems in understanding document images, especially in cases where training data is limited. We address the problem by proposing a novel finetuning method using reinforcement learning. Our approach treats the Information Extraction model as a policy network and uses policy gradient training to update the model to maximize combined reward functions that complement the traditional cross-entropy losses. Our experiments on four datasets using labels and expert feedback demonstrate that our finetuning mechanism consistently improves the performance of a state-of-the-art information extractor, especially in the small training data regime.

Meta reinforcement learning (Meta-RL) is an approach wherein the experience gained from solving a variety of tasks is distilled into a meta-policy. The meta-policy, when adapted over only a small (or just a single) number of steps, is able to perform near-optimally on a new, related task. However, a major challenge to adopting this approach to solve real-world problems is that they are often associated with sparse reward functions that only indicate whether a task is completed partially or fully. We consider the situation where some data, possibly generated by a sub-optimal agent, is available for each task. We then develop a class of algorithms entitled Enhanced Meta-RL using Demonstrations (EMRLD) that exploit this information even if sub-optimal to obtain guidance during training. We show how EMRLD jointly utilizes RL and supervised learning over the offline data to generate a meta-policy that demonstrates monotone performance improvements. We also develop a warm started variant called EMRLD-WS that is particularly efficient for sub-optimal demonstration data. Finally, we show that our EMRLD algorithms significantly outperform existing approaches in a variety of sparse reward environments, including that of a mobile robot.

Robotic manipulation and control has increased in importance in recent years. However, state of the art techniques still have limitations when required to operate in real world applications. This paper explores Hindsight Experience Replay both in simulated and real environments, highlighting its weaknesses and proposing reinforcement-learning based alternatives based on reward and goal shaping. Additionally, several research questions are identified along with potential research directions that could be explored to tackle those questions.

Non-prehensile pushing actions have the potential to singulate a target object from its surrounding clutter in order to facilitate the robotic grasping of the target. To address this problem we utilize a heuristic rule that moves the target object towards the workspace's empty space and demonstrate that this simple heuristic rule achieves singulation. We incorporate this effective heuristic rule to the reward in order to train more efficiently reinforcement learning (RL) agents for singulation. Simulation experiments demonstrate that this insight increases performance. Finally, our results show that the RL-based policy implicitly learns something similar to one of the used heuristics in terms of decision making. Qualitative results, code, pre-trained models and simulation environments are available at https://github.com/robot-clutter/improved_rl.