In the course of obtaining the results just described, various network topolo- gies and incremental evolution approaches were investigated (for details see [29]). Without incorporating some domain knowledge into the evolutionary process, we were unable to evolve successful controllers. Domain knowledge was introduced in the form of the network topology, which neglects any coupling between the
Fig. 3. Trajectory of a modular network controller after complet- ing 1100 timesteps of the waypoint
task. Fig. 4. Plot of the commanded speed vs the
real helicopter speed for the best controller
evolved with the velocity task.
lateral, longitudinal, and vertical axes. Evolving the yaw controller first was also crucial for the evolutionary process, confirming the findings of other researchers about the nature and benefits of incremental evolution.
5 Future work
In the immediate future we will implement and evaluate the approach of data collection, state estimation, system identification, and controller design on our model helicopter.
The first step will concentrate on the validation of the unscented Kalman filtering approach; the maximum update frequency and also the numerical ro- bustness need to be determined. The performance of the filter algorithm in terms of noise and drift also needs to be tested to ensure that the data will be adequate for control.
Model identification based on recorded flight data will constitute the next step. By its very nature the system identification technique will provide us with a quantitative estimation of the error between the simulated and real trajectory. We expect that the simulator will not be able to predict the trajectory of the real helicopter for more then a short period of time, due to the accumulation of error. However, the dynamic response of the model to the control input, which is what is needed to evolve a controller, will always resemble that of the real helicopter.
Artificial evolution will then be applied to produce controllers tailored to our helicopter. Several controllers chosen from those with good fitness will than be evaluated directly on the real helicopter.
Finally, a controller will be implemented on board the helicopter and the sensor to motor action loop will be closed, allowing us to test autonomous flight. The work will then proceed with the investigation of strategies for achieving flocking; these will initially be based on the classical rules of cohesion, separation,
6 Concluding remarks
The work presented here is clearly still in its early stages, but is following a clear path supported by existing research findings. The results achieved in the simulation and testing carried out so far are encouraging; we recognise however that porting the results obtained in simulation to a real system is very often problematical.
References
1. Reynolds, C.: Flocks, herds, and schools: A distributed behavioral model. In: Proceedings of the Conference on Computer Graphics (SIGGRAPH). Volume 21:4. (1987) 25–34
2. Mataric, M.: Interaction and intelligent behavior. PhD thesis, Massachusetts Institute of Technology, Cambridge, MA, USA (1995)
3. Kelly, I., Keating, D.: Flocking by the fusion of sonar and active infrared sensors on physical autonomous mobile robots. In: Proceedings of The Third Int. Conf. on Mechatronics and Machine Vision in Practice. Volume 1. (1996) 1–4
4. Welsby, J., Melhuish, C.: Autonomous minimalist following in three dimensions: A study with small-scale dirigibles. In: Proceedings of Towards Intelligent Mobile Robots Manchster. (2001) 无人机英文文献和中文翻译(8):http://www.751com.cn/fanyi/lunwen_77356.html