We propose a novel approach to addressing two fundamental challenges in Model-based Reinforcement Learning (MBRL): the computational expense of repeatedly finding a good policy in the learned model, and the objective mismatch between model fitting and policy computation. Our "lazy" method leverages a novel unified objective, Performance Difference via Advantage in Model, to capture the performance difference between the learned policy and expert policy under the true dynamics. This objective demonstrates that optimizing the expected policy advantage in the learned model under an exploration distribution is sufficient for policy computation, resulting in a significant boost in computational efficiency compared to traditional planning methods. Additionally, the unified objective uses a value moment matching term for model fitting, which is aligned with the model’s usage during policy computation. We present two no-regret algorithms to optimize the proposed objective, and demonstrate their statistical and computational gains compared to existing MBRL methods through simulated benchmarks.

Search-based techniques have shown great success in motion planning problems such as robotic navigation by discretizing the state space and precomputing motion primitives. However in domains with complex dynamic constraints, constructing motion primitives in a discretized state space is non-trivial. This requires operating in continuous space which can be challenging for search-based planners as they can get stuck in local minima regions. Previous work [1] on planning in continuous spaces introduced soft duplicate detection which requires search to compute the duplicity of a state with respect to previously seen states to avoid exploring states that are likely to be duplicates, especially in local minima regions. They propose a simple metric utilizing the Euclidean distance between states, and proximity to obstacles to compute the duplicity. In this paper, we improve upon this metric by introducing a kinodynamically informed metric, subtree overlap, between two states as the similarity between their successors that can be reached within a fixed time horizon using kinodynamic motion primitives. This captures the intuition that, due to robot dynamics, duplicate states can be far in Euclidean distance and result in very similar successor states, while non-duplicate states can be close and result in widely different successors. Our approach computes the new metric offline for a given robot dynamics, and stores the subtree overlap value for all possible relative state configurations. During search, the planner uses these precomputed values to speed up duplicity computation, and achieves fast planning times in continuous spaces in addition to completeness and sub-optimality guarantees. Empirically, we show that our improved metric for soft duplicity detection in search-based planning outperforms previous approaches in terms of planning time, by a factor of 1.5 to 2× on 3D and 5D planning domains with highly constrained dynamics.

Given access to accurate dynamical models, modern planning approaches are effective in computing feasible and optimal plans for repetitive robotic tasks. However, it is difficult to model the true dynamics of the real world before execution, especially for tasks requiring interactions with objects whose parameters are unknown. A recent planning approach, CMAX, tackles this problem by adapting the planner online during execution to bias the resulting plans away from inaccurately modeled regions. CMAX, while being provably guaranteed to reach the goal, requires strong assumptions on the accuracy of the model used for planning and fails to improve the quality of the solution over repetitions of the same task. In this paper we propose CMAX++, an approach that leverages real-world experience to improve the quality of resulting plans over successive repetitions of a robotic task. CMAX++ achieves this by integrating model-free learning using acquired experience with model-based planning using the potentially inaccurate model. We provide provable guarantees on the completeness and asymptotic convergence of CMAX++ to the optimal path cost as the number of repetitions increases. CMAX++ is also shown to outperform baselines in simulated robotic tasks including 3D mobile robot navigation where the track friction is incorrectly modeled, and a 7D pick-and-place task where the mass of the object is unknown leading to discrepancy between true and modeled dynamics.

We consider the task of autonomously unloading boxes from trucks using an industrial manipulator robot. There are multiple challenges that arise: (1) real-time motion planning for a complex robotic system carrying two articulated mechanisms, an arm and a scooper, (2) decision-making in terms of what action to execute next given imperfect information about boxes such as their masses, (3) accounting for the sequential nature of the problem where current actions affect future state of the boxes, and (4) real-time execution that interleaves high-level decision-making with lower level motion planning. In this work, we propose a planning, learning, and reasoning framework to tackle these challenges, and describe its components including motion planning, belief space planning for offline learning, online decision-making based on offline learning, and an execution module to combine decision-making with motion planning. We analyze the performance of the framework on real-world scenarios. In particular, motion planning and execution modules are evaluated in simulation and on a real robot, while offline learning and online decision-making are evaluated in simulated real-world scenarios.

Black-box optimizers that explore in parameter space have often been shown to outperform more sophisticated action space exploration methods developed specifically for the reinforcement learning problem. We examine these black-box methods closely to identify situations in which they are worse than action space exploration methods and those in which they are superior. Through simple theoretical analyses, we prove that complexity of exploration in parameter space depends on the dimensionality of parameter space, while complexity of exploration in action space depends on both the dimensionality of action space and horizon length. This is also demonstrated empirically by comparing simple exploration methods on several model problems, including Contextual Bandit, Linear Regression and Reinforcement Learning in continuous control.

We study Imitation Learning (IL) from Observations alone (ILFO) in large-scale MDPs. While most IL algorithms rely on an expert to directly provide actions to the learner, in this setting the expert only supplies sequences of observations. We design a new model-free algorithm for ILFO, Forward Adversarial Imitation Learning (FAIL), which learns a sequence of time-dependent policies by minimizing an Integral Probability Metric between the observation distributions of the expert policy and the learner. FAIL provably learns a near-optimal policy with a number of samples that is polynomial in all relevant parameters but independent of the number of unique observations. The resulting theory extends the domain of provably sample efficient learning algorithms beyond existing results that typically only consider tabular RL settings or settings that require access to a near-optimal reset distribution. We also demonstrate the efficacy ofFAIL on multiple OpenAI Gym control tasks.

Robots that navigate through human crowds need to be able to plan safe, efficient, and human predictable trajectories. This is a particularly challenging problem as it requires the robot to predict future human trajectories within a crowd where everyone implicitly cooperates with each other to avoid collisions. Previous approaches to human trajectory prediction have modeled the interactions between humans as a function of proximity. However, that is not necessarily true as some people in our immediate vicinity moving in the same direction might not be as important as other people that are further away, but that might collide with us in the future. In this work, we propose Social Attention, a novel trajectory prediction model that captures the relative importance of each person when navigating in the crowd, irrespective of their proximity. We demonstrate the performance of our method against a state-of-the-art approach on two publicly available crowd datasets and analyze the trained attention model to gain a better understanding of which surrounding agents humans attend to, when navigating in a crowd.

For robots to be a part of our daily life, they need to be able to navigate among crowds not only safely but also in a socially compliant fashion. This is a challenging problem because humans tend to navigate by implicitly cooperating with one another to avoid collisions, while heading toward their respective destinations. Previous approaches have used handcrafted functions based on proximity to model human-human and human-robot interactions. However, these approaches can only model simple interactions and fail to generalize for complex crowded settings. In this paper, we develop an approach that models the joint distribution over future trajectories of all interacting agents in the crowd, through a local interaction model that we train using real human trajectory data. The interaction model infers the velocity of each agent based on the spatial orientation of other agents in his vicinity. During prediction, our approach infers the goal of the agent from its past trajectory and uses the learned model to predict its future trajectory. We demonstrate the performance of our method against a state-of-the-art approach on a public dataset and show that our model outperforms when predicting future trajectories for longer horizons.

Path planning in the presence of dynamic obstacles is a challenging problem due to the added time dimension in search space. In approaches that ignore the time dimension and treat dynamic obstacles as static, frequent re-planning is unavoidable as the obstacles move, and their solutions are generally sub-optimal and can be incomplete. To achieve both optimality and completeness, it is necessary to consider the time dimension during planning. The notion of adaptive dimensionality has been successfully used in high-dimensional motion planning such as manipulation of robot arms, but has not been used in the context of path planning in dynamic environments. In this paper, we apply the idea of adaptive dimensionality to speed up path planning in dynamic environments for a robot with no assumptions on its dynamic model. Specifically, our approach considers the time dimension only in those regions of the environment where a potential collision may occur, and plans in a low-dimensional state-space elsewhere. We show that our approach is complete and is guaranteed to find a solution, if one exists, within a cost sub-optimality bound. We experimentally validate our method on the problem of 3D vehicle navigation (x, y, heading) in dynamic environments. Our results show that the presented approach achieves substantial speedups in planning time over 4D heuristic-based A*, especially when the resulting plan deviates significantly from the one suggested by the heuristic.