3D Mobility Learning and Regression of Articulated, Tracked Robotic Vehicles by Physics-based Optimization

Panagiotis Papadakis, Fiora PirriVRIPHYS 2012 · DOI

TL;DR

This paper introduces a physics-based simulation framework to estimate the stable pose and assess the 3D traversability of articulated, tracked robots on rough terrain. By unifying robot and terrain geometry with physics, it enables robots to learn and predict their mobility, crucial for complex environments like urban search and rescue.

What Problem Does This Paper Address?

Navigating complex, unstructured 3D terrain is a significant challenge for mobile robots, especially those designed for environments like urban search and rescue (USAR). Existing approaches often simplify the problem by treating robots as basic geometric shapes or focusing on 2D terrain representations. This paper addresses the gap by proposing a method to accurately assess the static traversability of articulated, tracked robots on real 3D terrain. This involves not only understanding if a robot can physically be in a certain pose but also quantifying how optimal that pose is, considering the robot's intrinsic characteristics and its interaction with the terrain. The goal is to enable robots to learn their mobility skills and adapt their configuration for safer and more efficient traversal.

Methods

The core of the methodology is a physics-based simulation pipeline that operates in three sequential stages:

  1. Stable State Estimation (Chassis Elevation and Orientation): This stage estimates the robot's stable 3D position and orientation (roll, pitch, yaw) when placed on a given terrain patch. It uses a physics-based simulation that accounts for gravity, friction, and terrain properties (like softness and bounciness) to find the pose where the robot's linear and angular velocities converge to zero. This is achieved by iterating until the chassis reaches a stable equilibrium.
  2. Stable State Estimation (Articulating Bodies Rotations): Building on the stable chassis pose, this stage further refines the robot's stability by adjusting the rotations of its articulating components (e.g., flippers). The goal is to maximize contact with the terrain to increase stability and reduce the risk of tipping over, while respecting the robot's kinematic constraints.
  3. Static 3D Traversability Cost Assessment: Once the stable pose is determined, this stage quantifies the "optimality" of that state using a set of cost criteria:
  4. Ground Clearance (G): An improved measure using the directed Hausdorff distance between the robot's base and the terrain, accounting for the detailed geometry of both.
  5. Robot Orientation (O): Considers not only roll and pitch but also yaw drift that can occur during the stabilization process, penalizing deviations from the intended path.
  6. Angle Stability (A): Based on the force-angle stability margin, this criterion penalizes the tip-over angle as it approaches zero, providing a scale-invariant measure.

The entire framework relies on having a 3D model of the robot and a 3D representation of the terrain (e.g., a DEM or polygonal surface derived from point clouds).

Datasets

Not discussed in the paper. The paper focuses on a simulated model of the TALOS robot and uses synthetic terrain data generated for experiments.

Key Insights and Ideas

  • Unifying Physics and Geometry for Traversability: The paper demonstrates that a robust assessment of traversability requires more than just geometric analysis; it necessitates incorporating physics-based simulation to understand the dynamic interaction between the robot and the terrain.
  • Importance of Articulation in Stability: For articulated robots, simply finding a stable chassis pose isn't enough. Adjusting the articulating components is crucial for maximizing stability and adapting to uneven surfaces.
  • Refined Cost Metrics: The proposed cost metrics (especially the directed Hausdorff distance for ground clearance and the inclusion of yaw drift for orientation) offer a more nuanced and accurate assessment of traversability compared to traditional methods.
  • Real-time Applicability: The physics-based simulation, despite its complexity, is shown to be computationally efficient enough for real-time traversability assessment, a critical requirement for onboard robotic systems.

Key Results

  • The proposed methodology can efficiently estimate the stable pose of an articulated, tracked robot on 3D terrain.
  • The traversability cost assessment is performed in real-time, with computation times for stabilization and cost calculation being relatively constant and independent of terrain complexity.
  • Experiments with a simulated TALOS robot show that the traversability costs (G, O, A) evolve predictably with increasing terrain complexity (slope and roughness).
  • The method successfully visualizes stabilized robot states on various challenging USAR-relevant terrains, including inclined planes and steps.

Limitations and Future Work

  • Real-time Motion Trajectory Simulation: While static traversability assessment is real-time, simulating the motion trajectory of the robot using its tracks along a sequence of poses still remains a challenge due to computational constraints.
  • Dynamic Environments and Uncertainty: The current framework focuses on static traversability. Future work could explore dynamic environments, dealing with uncertainty in terrain perception, and handling execution failures.
  • Learning and Regression: The paper mentions "mobility regression" but the detailed implementation and results of this aspect are not fully elaborated. Further work could focus on more sophisticated learning and regression techniques to predict mobility based on terrain characteristics.
  • Sensor Integration: The paper assumes readily available 3D terrain models. Integrating this with real-time sensor data (like LiDAR or stereo vision) for dynamic terrain mapping and adaptation would be a crucial next step for practical deployment.

Summary updated

BibTeX
@inproceedings{papadakis2012d,
  author = {Panagiotis Papadakis and Fiora Pirri},
  title = {{3D Mobility Learning and Regression of Articulated, Tracked Robotic Vehicles by Physics-based Optimization}},
  booktitle = {VRIPHYS},
  year = {2012},
  doi = {10.2312/pe/vriphys/vriphys12/147-156},
}
Figure