"Mana" rover

The work done here was a preliminary analysis of rover locomotion and diagnostics for various terrain types (grass, rocks, asphalt) and different locomotion classes (nominal, rough, failure). This was a preparatory study supposed to be further extended into a PhD thesis “Agile Locomotion for a Planetary Rover”, in collaboration with Astrium Astrium and ESA. Unfortunately, Astrium went under restructuring and was unable to finance it.

The motivation lies in the fact that the chassis of planetary rovers is conceived to endow them with high locomotion capabilities, however the classical navigation loops considered for now (and in the near future, e.g. for the Exomars rover) are defined so that even small obstacles must be circumnavigated. There is however a big interest in empowering such rovers with obstacle-crossing capabilities, from an operational point of view on the one hand, but also from the scientific point (collecting more interesting samples) on the other hand, as this remains an open challenging problem. Planetary exploration is just an example application, the work would also pertain to any robot with advanced locomotion capabilities, such as in search and rescue applications for instance.

CHALLENGES

  • Such a locomotion ability would require a thorough re-visit of the navigation loop, with a very strong focus on locomotion control: when trying to climb a pile of rocks for instance, the 3D model built is of little help as many areas remain occluded (due to the terrain geometry and the angle from which range data is acquired), and also because when driving on such areas, the ground below the wheels is "moving“ (rocks are rolling, wheels are slipping, etc.)
  • One of the most difficult issues is that the knowledge of the wheel/soil interactions is very poor. Hence, a challenge is to derive such knowledge as much as possible from all the parameters that can be sensed on-board the rover (from wheel current consumption to overall position and speed, via all the encoders information, terrain analysis or even sound).

A “diagnosis” capability is also a part of the problem. A “classic” example is when the Opportunity rover got stuck in a loose sand dune, back in April 2005, it took 5 weeks for NASA engineers to free the rover. There were many options to detect and prevent this (e.g. wheel slippage detection, even without comparison with visual odometry), the problem was that the loose sand had no visual print, and at that time they had already traversed a few kilometers on this monotonic environment, looking forward to reach Victoria crater. This actually pertains more to locomotion diagnostic (or monitoring), but still is something which is part of "agile locomotion".

Work done at LAAS consisted of gathering and analysing the data using the “Mana” rover

MANA Rover

  • Inputs from sensors gathered
    • Velocity, torque, IMU...
  • Labeling by expert annotator
    • 3 classes (normal, rough and fault locomotion)
  • Machine Learning modelling methods explored
    • Conditional Random Fields (CRF), Hidden Markov Models (HMM), Naive Bayes…

PRELIMINARY RESULTS

Due to the limited amount of time, the analysis of results was not completed. However, based on the obtained feature-label distributions the problem seems to be complex, but interesting for further examination.

Label distribution in the space of front - back wheel velocity and torque difference:

Label distribution in the space of left – right wheel velocity and torque difference: