Multi-Sample Long Range Path Planning under Sensing Uncertainty for Off-Road Autonomous Driving

TL;DR

The method, Dynamic Replanning via Evaluating and Aggregating Multiple Samples (DREAMS), outperforms other determinization-based approaches in terms of combined traversal time and collision cost and aims to be robust to uncertainty over potential worlds while still achieving the efficiency benefits of determinization.

Abstract

We focus on the problem of long-range dynamic replanning for off-road autonomous vehicles, where a robot plans paths through a previously unobserved environment while continuously receiving noisy local observations. An effective approach for planning under sensing uncertainty is determinization, where one converts a stochastic world into a deterministic one and plans under this simplification. This makes the planning problem tractable, but the cost of following the planned path in the real world may be different than in the determinized world. This causes collisions if the determinized world optimistically ignores obstacles, or causes unnecessarily long routes if the determinized world pessimistically imagines more obstacles. We aim to be robust to uncertainty over potential worlds while still achieving the efficiency benefits of determinization. We evaluate algorithms for dynamic replanning on a large real-world dataset of challenging long-range planning problems from the DARPA RACER program. Our method, Dynamic Replanning via Evaluating and Aggregating Multiple Samples (DREAMS), outperforms other determinization-based approaches in terms of combined traversal time and collision cost. https://sites.google.com/cs.washington.edu/dreams/

Figures (7)

• Figure 1: An autonomous off-road vehicle's long-range planner needs to decide the best way up a hill, given blind spots and imperfect sensing.
• Figure 2: Overview of DREAMS. Sample & Plan: Sample many worlds from the posterior distribution, and plan the optimal path on a subsample of worlds ($\phi_1$, $\phi_{10}$, $\phi_{50}$ above). Evaluate: Evaluate the cost of each resulting plan against the full set of sampled worlds. Aggregate: Aggregate the resulting cost distribution with a summary statistic (e.g., mean or CVaR) Select: Select the plan with minimal aggregated cost.
• Figure 3: The sensor noise levels used in testing: $\eta_{\mathrm{low}} = 10^{-4}$, $\eta_{\mathrm{med}} = 10^{-3}$, $\eta_{\mathrm{high}} = 10^{-2}$. Our simplified noise model defines the probability of receiving a correct observation for a query point distance $d$ away as $\max ( \exp(-\eta d^2), p_{\mathrm{min}})$. The minimum probability threshold $p_{\mathrm{min}}$ is set to 0.6 to provide some signal at the edge of the robot's observation range; note that the minimum possible value of $p_{\mathrm{min}}$ for binary occupancy is 0.5 (pure noise). With these parameters, the robot receives approximately 96%, 72%, and 61% correctly observed pixels per observation.
• Figure 4: Traversed paths of each algorithm (blue edges) on two example worlds from each of the Forest and Desert datasets, with the same world in each set of four. Each algorithm receives observations with high noise, and is penalized with a collision factor of $\alpha=10$. With high noise, DRPS frequently backtracks and changes direction while Sampled A* incurs many collisions (red edges). Both DREAMS variants follow more direct paths without collisions.
• Figure 5: Qualitative comparison of each approach, given the exact same sampled worlds and paths. Robot (blue), proposed paths (light orange), accepted path (bright orange). Top: All plans except DRPS find a path through the gap. DRPS happened to sample a world that did not fit through the gap, producing a longer route. Bottom: All plans except Sampled A* reverse from a likely obstacle in front of the robot. Sampled A* does not explicitly consider the cost of collision and accepts a path going through the obstacle. Right: Looking at the heatmap, areas where more plans overlap are hotter. As there is little overlap besides the start position, Sampled A* has less signal to choose the most likely plan; its decision is almost a uniform random sample.