In Multi-Policy Decision-Making (MPDM), many computationally-expensive forward simulations are performed in order to predict the performance of a set of candidate policies. In risk-aware formulations of MPDM, only the worst outcomes affect the decision making process, and efficiently finding these influential outcomes becomes the core challenge. Recently, stochastic gradient optimization algorithms, using a heuristic function, were shown to be significantly superior to random sampling. In this paper, we show that accurate gradients can be computed - even through a complex forward simulation - using approaches similar to those in deep networks. We show that our proposed approach finds influential outcomes more reliably, and is faster than earlier methods, allowing us to evaluate more policies while simultaneously eliminating the need to design an easily-differentiable heuristic function. We demonstrate significant performance improvements in simulation as well as on a real robot platform navigating a highly dynamic environment.

Risk-aware Multi-Policy Decision Making (MPDM)is a powerful framework for reliable navigation in a dynamic social environment where rather than evaluating individual trajectories, a “library” of policies (reactive controllers)is evaluated by anticipating potentially dangerous future outcomes using an on-line forward roll-out process. There is a core tension in Multi-Policy Decision Making (MPDM)systems - it is desirable to add more policies to the system for flexibility in finding good policies, however, this increases computational cost. As a result, MPDM was limited to small (perhaps 5-10)discrete policies - a significant performance bottleneck. In this paper, we radically enhance the expressivity of MPDM by allowing policies to have continuous-valued parameters, while simultaneously satisfying real-time constraints by quickly discovering promising policy parameters through a novel iterative gradient-based algorithm. Our evaluation includes results from extensive simulation and real-world experiments in semi-crowded environments.

In the Multi-Policy Decision Making (MPDM) framework, a robot's policy is elected by sampling from the distribution of current states, predicting future outcomes through forward simulation, and selecting the policy with the best expected performance. Electing the best plan depends on sampling initial conditions with influential (very high costs) outcomes. Discovering these configurations through random sampling may require drawing many samples, which becomes a performance bottleneck. In this paper, we describe a risk-aware approach which augments this sampling with an optimization process that helps discover those influential outcomes. We describe how we overcome several practical difficulties with this approach, and demonstrate significant performance improvements on a real robot platform navigating a semi-crowded, highly dynamic environment.

In dynamic environments crowded with people, robot motion planning becomes difficult due to the complex and tightly-coupled interactions between agents. Trajectory planning methods, supported by models of typical human behavior and personal space, often produce reasonable behavior. However, they do not account for the future closed-loop interactions of other agents with the trajectory being constructed. As a consequence, the trajectories are unable to anticipate cooperative interactions (such as a human yielding), or adverse interactions (such as the robot blocking the way). In this paper, we propose a new method for navigation amongst pedestrians in which the trajectory of the robot is not explicitly planned, but instead, a planning process selects one of a set of closed-loop behaviors whose utility can be predicted through forward simulation. In particular, we extend Multi-Policy Decision Making (MPDM) [1] to this domain using the closed-loop behaviors Go-Solo, Follow-other, and Stop. By dynamically switching between these policies, we show that we can improve the performance of the robot as measured by utility functions that reward task completion and penalize inconvenience to other agents. Our evaluation includes extensive results in simulation and real-world experiments.