End-to-End Statistical World Representations – Terrain Mapping to Traversability Analysis


  1. Broten, G. S.
  2. Collier, J. A.
  3. Mackay, D. J.
  4. Monckton, S. P.
  5. Digney, B. L.
Corporate Authors
Defence R&D Canada - Suffield, Ralston ALTA (CAN)
Robot navigation requires reliable perception that generates an appropriate world representation. This is especially true for outdoor, unstructured or semi-structured environments where impediments to traversal are more complex than simple insurmountable obstacles. Such environments include negative features, such as ditches, as well as positive features in the form of sloping rises, that may constitute obstacles to traversal. These features must be captured by the world representation in a manner that properly handles the uncertainty associated with range data and the vehicle pose. An analysis of this representation can then extract traversable and non-traversable regions. Finally, this processing must proceed in near real-time in order to allow the robot to travel at reasonable speeds. This paper presents a statistical approach that starts with a 2.5D terrain map, follows through to the traversability analysis, and continues to the ancillary global terrain and traversability representations. At all stages the data’s statistical relevance is carried forward and incorporated into the analysis. This technique has been verified using simulations and heavily exercised on a physical robot under real world conditions. These experiments have revealed that the proposed approach performs well, producing both terrain and traversability maps that adequately portray their environment.

Il y a un résumé en français ici.

self-supervised;on-line;real-time;learning;adaptive;inferred geometry;terrain mapping
Report Number
DRDC-SUFFIELD-TR-2013-060 — Technical Report
Date of publication
01 Nov 2013
Number of Pages
Electronic Document(PDF)

Permanent link

Document 1 of 1

Date modified: