Motion generation for a rover on rough terrains

David Bonnafous, Simon Lacroix, Thierry Siméon

TL;DR

This paper presents a novel algorithm for generating safe and efficient motion trajectories for an articulated rover navigating challenging, rough terrains. By explicitly considering the rover's chassis geometry and kinematic constraints, the algorithm enables robust navigation in previously unexplorable environments.

What Problem Does This Paper Address?

Autonomous navigation in unstructured and rough environments presents a significant challenge for mobile robots. Unlike smooth, predictable surfaces, rough terrains introduce complexities in perception, trajectory planning, and execution. Specifically, accurately modeling the terrain's geometry, accounting for the robot's own physical limitations (like its articulated chassis), and ensuring the safe execution of planned movements are critical. Existing approaches often struggle with the dynamic interactions between an articulated robot and uneven ground, especially when dealing with uncertainties in terrain data. This paper focuses on the crucial aspect of determining feasible and safe motions for an articulated rover on such challenging terrains.

Methods

The core of the proposed solution is a motion generation algorithm that evaluates a set of elementary trajectories, specifically circle arcs, on a continuously updated digital elevation map (DEM). The algorithm's novelty lies in its explicit consideration of the rover's chassis geometry and kinematic constraints.

Here's a breakdown of the key components:

  • Digital Elevation Map (DEM) Building: The system builds a DEM from stereo vision data. To handle occlusions and data resolution issues common in rough terrain, the DEM is continuously updated as the robot moves, relying on visual odometry for precise 3D position estimation.
  • Chassis Pose and Configuration Determination (placeRobot function): This function is crucial for predicting the rover's internal configuration (angles of its passive articulations) and its pose on the DEM for any given position. It iteratively determines the placement of each axle by considering wheel-terrain interactions and the robot's geometry.
  • Motion Control: The rover "Lama" uses a skid-steering scheme for motion control. The system can handle linear and angular velocity commands by controlling the wheel speeds.
  • Motion Generation Algorithm: This is the heart of the system. It evaluates candidate circle arcs by discretizing them into a series of configurations. For each configuration, it calculates a "danger" score based on:
  • Internal Chassis Constraints: Ensuring the rover's passive articulations remain within their physical limits.
  • Stability Constraints: Preventing excessive pitch and roll angles on each axle, which could lead to instability.
  • Unknown Terrain Areas: Accounting for regions in the DEM that are not perceived due to obstacles or data limitations. The "danger" is normalized into a constraint set. The "cost" of a motion (a circle arc) integrates this danger score with the distance traveled and the change in the robot's internal configuration. This cost is then used with an A* search algorithm to select the optimal trajectory. The heuristic for the A* search is the Dubbins path distance to the goal.

Datasets

Not discussed in the paper. The paper focuses on an algorithm tested on a real robot ("Lama") in various real-world terrain conditions.

Key Insights and Ideas

  • Explicit Geometric Constraints are Key: The paper strongly emphasizes that for articulated robots on rough terrain, simply having a DEM is insufficient. Explicitly modeling and checking the robot's internal configuration and stability constraints at every potential step is paramount for safe navigation.
  • "Danger" as a Unified Metric: By combining various constraints (chassis articulation, stability, unknown terrain) into a single "danger" score, the algorithm provides a unified way to assess the risk associated with a particular robot configuration.
  • Circle Arcs as Feasible Elementary Motions: The choice of circle arcs as elementary trajectories is practical because they are directly controllable by the rover's skid-steering mechanism, simplifying the motion execution layer.
  • "Look Ahead" Behavior: Discretizing arcs and evaluating configurations up to a certain distance ahead of the robot allows for a "look ahead" behavior, helping to avoid immediate obstacles and plan more holistically.

Key Results

The algorithm was successfully integrated on board the robot "Lama" and demonstrated its effectiveness in numerous experiments on various rough terrains.

  • Robust Navigation: The rover was able to autonomously reach goals tens of meters away in initially unknown terrains.
  • Real-time Performance: The motion generation algorithm achieved a mean execution time of approximately 300ms per step, making it suitable for real-time navigation.
  • Experimental Validation: Figure 10 shows a comparison between predicted and measured chassis configuration angles, demonstrating the accuracy of the placeRobot function. Figures 11 and 12 showcase successful trajectories, including navigating a narrow corridor and climbing a hill.

Limitations and Future Work

  • Dynamic Issues: The authors acknowledge that considering realistic dynamic issues (e.g., forces, accelerations, soil mechanics) is a significant open problem. While pioneer work exists, its implementation on real rovers requires further research and validation.
  • Uncertainty in Terrain Models: While the DEM is continuously updated, uncertainties in the 3D data and the resulting DEM can still pose challenges.
  • Future Work: The paper suggests that further work is needed to incorporate dynamic considerations into the motion planning process.

Summary updated

Abstract

International audience
BibTeX
@inproceedings{bonnafous2001motion,
  author = {David Bonnafous and Simon Lacroix and Thierry Siméon},
  title = {{Motion generation for a rover on rough terrains}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2001},
  doi = {10.1109/iros.2001.976264},
}
Figure