This research in long-range autonomous rover navigation seeks to enable efficient path planning and reliable navigation for single-command, one-kilometer autonomous rover traverse. It builds upon our previous work in global path planning
The scientific data return on Mars rover exploration missions is
fundamentally limited by the operational overhead associated with
energy collection, power management, communications and travel between
interesting sites. Time efficiency can be achieved through a
combination of detailed planning at low frequency on Earth and fast
response replanning on the rover, thereby enabling intelligent
decisions at a far greater bandwidth than possible under direct human
control.
Using our solar-powered rover Hyperion, we have demonstrated
preliminary results in long-range autonomous navigation with traverses
of 300-600 meters in a single command using 30-meter resolution
digital elevation models and odometric position estimation (meaning
without GPS). We have achieved one instance of single-command
1-kilometer autonomous traverse in the
Chilean Desert in April
2003.
We are working to enable long traverses that are
reliable and efficient. By reliable, we mean that the rover minimizes
the risk that it will be lost, damaged, or require remote
operation. By efficient, we mean that the rover minimizes critical
resources to accomplish its objectives.
Our present path planning algorithms reason about distance, time,
energy, and risk, but produce routes that are optimal only in
discretized planning space. We extending the algorithms to the
continuous realm to find the truly optimal routes, thus improving both
efficiency and reliability (if risk is minimized in the plan). At
present, we produce routes based on digital elevation models. As the
rover drives, it updates its map with data from onboard sensors and
re-plans as needed. This makes limited use of the sensor information.
We propose to update all related areas of the map so that the benefit
is global, not just local. The newly planned routes are based on a
more accurate map, and thus are more reliable and efficient.
Present navigation algorithms focus on the near-field, leaving a gap
between the maximum range analyzed by the rover (< 10m) and the
minimum resolution of orbital models (>30m resolution). This myopic
view is prone to driving into super-obstacles like embankments,
drainages, craters, hummocks, and dunes. We propose to extend the
perception analysis to the far-field and modeling to multiple scales
to enable planning for these terrain features.
For information contact: Tony Stentz or: David Wettergreen
This research is supported by NASA under contract from JPL.