Landmark-based bike motion planning

In this work a very robust path-following bike controller was developed, as well as a number of strategies for laying down "realistic" guide paths (i.e., obstacle-hugging, similar to how a human would ride). Again, this planner is intended for local motion planning, with the global path strategy being provided by an external source (e.g., user, kinematic RRT or PRM, etc.)

The heart of the path-following controller lies in the following observations:

• for a particular bike lean, there exists a steering angle ("nominal") whose centrifugal force will precisely counteract the pull of gravity in that pose
• by steering tighter or looser than the nominal angle (difference = "oversteer"), we can effect a predictable acceleration in the bike's lean angle; this relationship can be mapped out experimentally (upper-left image)
• path following can be achieved by presenting a target to the bike some specific constant distance ahead on the guide path; the bike's configuration and relative target position map to a unique desired curved path, and this curvature in turn maps to a desired lean angle
• by setting up a PD controller between the current and desired lean angles one can simultaneously achieve robust balancing and path following

The bottom left image shows a spline-based guide-path for road navigation, while on the right we show various solutions for less structured environment. In these the coarse line is the global strategy (input), while the closer hugging line (cyan) is the computed guide-path.