Activities

No upcoming events

A semi-autonomous wheelchair as a problem of evolutionary Algorithms

Participants

• Nico Lüdike
• Matthias Sterz

Project Supervisors

• Benjamin Inden
• Julia Fröhlich

Abstract

• Semi-autonomous wheelchairs give people with disabilities the chance to navigate in their environment, even if their interaction abilities are severely limited.
• In most applications the system is provided a map of the environment, often with predefined paths.
• To navigate more freely in unknown surroundings we tried to setup a system that learns pathfinding and obstacle avoidance by means of neuroevolution.
• To evaluate the quality of the evolved control system also subjective user impressions should be taken into account.

Motivation

Navigation is a strenuous activity for people, who are depending on a wheelchair. A system, where a semi-autonomic wheelchair drives a user to his destination, without colliding with objects, is beneficial. To garantee this we need a high robustness. Also it may be beneficial for the user to have some controll about the driving-system. These two aspects are what incite us to the project EVOLU - a pathfinding, obstacle avoidance and user influence system by means of neuroevolution.

Application Szenario

The wheelchair should be able to move the user to any location he\she wants by saying the target. For example you are in the living room, getting hungry. So you want to move to the kitchen to make yourself something to eat. So you just say it out loud and your wheelchair will start to move to the kitchen without any collision.

Objectives

The project goals are

• Implement two interchangeable simulators to evolve an artificial neuronal network that is able to control a semi-autonomous wheelchair.
• A 2D-based simulator should be rather fast computable, a 3D-physics-based simulator will provide a more realistic simulation.
• Of course safety issues are crucial in the wheelchair scenario. Consequently it should not fall over, crash into an obstacle or turn at high speed.

Description

we have implemented two simulators to use with the Scalable Neuroevolution Project (SNEP) Software. They calculate the behaviour of the wheelchair in a virtual environment and the resulting fitness of the particular individual. Both the map layout and neuroevolution factors can be assigned using ini-files. The simulators are designed to be interchangable; so a 2D-based evolved individual can be used as a starting point for further physics-based evolution. In addition to visualisation at a desktop PC, the system provides an interface to the Instant Reality framework which allows us to run the 3D-simulator with several vitual reality setups.

Results

• Two simulators have been successfully implemented and the evolution provided suitable control networks.
• The interface to Instant Reality enhances the possibility to rate the quality of and evolved control network.
• During development the fitness function was repeatedly improved.

There is a demonstration video (ogv, 2.1 Mb) for the 2D version and this Video (ogv, 17.8 Mb) for the 3D version.: The videos show snapshots of the ongoing evolution process. Up next are some results, how much generations the wheelchair needed, to learn to behave himself correctly in a predefined environment. In our simple example, the wheelchair starts on a fixed position on an appartment with four empty rooms and needs to reach one goal. The first graph show our results of the 2D-implementation, with only simple physics. The second graph shows the 3D-implementation with advanced physics (ODE):

Discussion and Conclusion

• Alhough we had to cope with the loss of one team member we managed to implement the system as planned. Nevertheless the goal of subjective influenced evolution could not be treated to the desired extent.

Outlook

The virtual reality interface provides the possibility to take subjective user impressions into account. This feature should be used to further improve the quality of the evolved control system. To achive this a user study in a CAVE Virtual Reality Environment is planned, while the fitness function will be further improved.