MTG: Mapless Trajectory Generator with Traversability Coverage for Outdoor Navigation

TL;DR

This work presents a novel learning-based trajectory generation algorithm for outdoor robot navigation designed for global planning using limited onboard robot perception in mapless environments while ensuring comprehensive coverage of all traversable directions.

Abstract

We present a novel learning-based trajectory generation algorithm for outdoor robot navigation. Our goal is to compute collision-free paths that also satisfy the environment-specific traversability constraints. Our approach is designed for global planning using limited onboard robot perception in mapless environments while ensuring comprehensive coverage of all traversable directions. Our formulation uses a Conditional Variational Autoencoder (CVAE) generative model that is enhanced with traversability constraints and an optimization formulation used for the coverage. We highlight the benefits of our approach over state-of-the-art trajectory generation approaches and demonstrate its performance in challenging and large outdoor environments, including around buildings, across intersections, along trails, and off-road terrain, using a Clearpath Husky and a Boston Dynamics Spot robot. In practice, our approach results in a 6% improvement in coverage of traversable areas and an 89% reduction in trajectory portions residing in non-traversable regions. Our video is here: https://youtu.be/3eJ2soAzXnU

Figures (10)

• Figure 1: Trajectory generation in a campus environment. The top view is the robot's trajectory (around 350m). The bottom row shows generated trajectories at the two locations corresponding to green and yellow boxes, respectively. The trajectory closest to the global target, shown at the top, will be chosen to provide waypoints for the local planner. The local planner drives the robot and avoids collisions liang2022adaptiveon. In this complex outdoor scenario, our MTG global navigation method can efficiently and safely generate trajectories for global navigation and cover most of the traversable areas, including the narrow pedestrian sidewalks.
• Figure 2: Overall Pipeline of MTG: The inputs are several consecutive frames of Lidar point clouds and velocities of the robot, $\epsilon$ is a Normal distribution, and $\mathbf{c}$ is the condition value. Green circles represent distribution of $\mathbf{z}_k$, $p(\mathbf{z}_k)$, where $k\in\left\{1,...,K\right\}$. The last part represents the decoders, and each decoder generates one trajectory.
• Figure 3: The quality results of CVAE cvae, DLOW yuan2020dlow, and MTG. The top row is from the robot’s view, and the bottom row is the birds-eye-view. White areas are non-traversable areas. The CVAE generates trajectories very similar to each other, while DLOW has a large diversity but is not good in terms of traversability. Our approach generates trajectories that cover mostly all traversable directions and lie on only traversable areas.
• Figure 4: Trajectory generation with different traversable areas; the blue area corresponds to the region visible to the robot, and the purple areas denote people.
• Figure 5: Trajectory Confidence: The top row shows the camera view of the generated trajectories and the bottom row shows the bird-eye-view of the trajectories. The Cyan color represents the obstacles detected from the middle channel of the 3D Lidar.