Uncertainty for mobile robots

Supervisors

• Alexis Scheuer, alexis.scheuer@loria.fr
• Francis Colas, francis.colas@inria.fr

Context

To work with mobile robots, it is necessary to build a model of their motion. This model can be used either to better estimate the motion already performed or to plan future actions. In the first case, the question is to know where the robot actually is. This is typically solved using a combination of the motion model of the robot and the observation models from its sensors (camera or lasers). In the second case, the question is to decide the next motions so as to reach a given goal. This is typically solved using a simulation of the various actions.

The motion of a robot can be described with various levels of complexity. Kinematic models, which consider velocities and trajectories, are often used. On top of that, one can add dynamical elements like balance or force and torque. Finally a more accurate description of the environment can be used to compute contact or friction forces.

Of course, a more complex model will require both more parameters and knowledge of the environment, as well as more computational resources for its solving. Moreover, whichever the model, its accuracy is always limited and a real robot will behave differently than modelled. Therefore, there is always uncertainty on the real motion of a robot.

Subject

The aim of this PhD proposal is to characterise and use motion uncertainty of a mobile robot so as to yield more precise and safer motion. Most state-of-the-art models ignore uncertainty, which limits their validity. For instance, Effati, Skonieczny, and Balkcom (2024) propose a method for skid-steer robots, but for hard floors only, as otherwise ground friction deviates too much the robot from the modelled motion.

The planned characterisation will consist in defining motion uncertainty models from real data. Classically, parametric models have been used since they allow for a compact representation and, in particular in the Gaussian case, with which computation complexity is reduced (Borja, Mirats Tur, and Gordillo 2009). However these models are more constrained and it is sometimes necessary to consider non parametric models (for instance with particles (Bishop 2006)). Finally, with a greater data availability, a lot of research domains have successfully leveraged machine learning (Murphy 2022). The aim is therefore to explore the trade-off between accuracy and complexity offered for those various approached for the modelling of robot motion.

The developed uncertainty models will then be exploited in three modules for mobile robot autonomy: localisation, path-planning, and trajectory tracking (when the latter uses a local planning method) (“Chapter 5: Robot Motion” 2005).

• For localisation, the uncertainty model should offer a better estimate of the actual pose reached by the robot following the commands sent to its motors. This improved estimation should, in turn, improve the sensor-based localisation of the robot.
• For global and local path-planning, the uncertainty model should provide a more complex and accurate idea of the consequences of the planned commands. In particular, commands with a significant probability to yield a collision should be better detected and avoided (in a better way than increasing safety margins).

A major point of attention of this work will be the complexity of the computations with the developed models. Probabilistic inference is generally expensive and approximate inference (or exact inference with approximated models) might be necessary to bound computation time. It would therefore be important to quantify both the approximation errors and the added-value compared with standard deterministic models. Finally, the various representations are diversely suited to specific kinds of computation and a different usage might require a different approach.

Application

Required qualifications

• MSc in computer science
• programming skills in Python and/or C++
• notions of probabilistic inference
• knowledge of kinematic robot models

Process

The application must be sent by mail to all supervisors as an archive (.zip or .tgz) containing the following documents:

• CV
• cover letter
• short description (maximum one page) of your master thesis (or equivalent), even if it is still in progress
• degree certificates and transcripts for bachelor and master (or the last 5 years)
• master thesis (or equivalent) if it is already completed and publications if any (it is not expected to have any); web links to these documents are preferable, if possible

In addition, one recommendation letter from the person who supervises(d) your master thesis (or research project or internship) should be sent directly by their author by mail (see above). Incomplete applications might be ignored. Should there be a problem, contact us.