Rigidity-Based Surface Recognition for a Domestic Legged Robot
Kertész, Csaba (2016)
Kertész, Csaba
2016
IEEE Robotics and Automation Letters 1 1
309-315
Informaatiotieteiden yksikkö - School of Information Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:uta-201611282684
https://urn.fi/URN:NBN:fi:uta-201611282684
Tiivistelmä
Although the infrared (IR) range and motor force sensors have been rarely applied to the surface recognition of mobile robots, they are fused in this paper with accelerometer
and ground contact force sensors to distinguish six indoor surface types. Their sensor values are affected by the crawling gait period, therefore, certain components of the fast Fourier transform over these data are included in the feature vectors as well as remarkable discriminative power is observed for the same scalar statistics of different sensing modalities. The machine learning aspects are analyzed with random forests (RF)
because of their stable performance and some inherent, beneficial properties for the model development process. The robustness is evaluated with unseen data after the model
accuracy is estimated with cross-validation (CV), and regardless whether a Sony ERS-7 walks barefoot or wears socks, the forests achieve 94% accuracy. This result outperforms the state of the art techniques for indoor surfaces in the literature and the
classification execution is real-time on the robot. The above mentioned model development process with RF is documented to create new models for other robots more quickly and efficiently.
and ground contact force sensors to distinguish six indoor surface types. Their sensor values are affected by the crawling gait period, therefore, certain components of the fast Fourier transform over these data are included in the feature vectors as well as remarkable discriminative power is observed for the same scalar statistics of different sensing modalities. The machine learning aspects are analyzed with random forests (RF)
because of their stable performance and some inherent, beneficial properties for the model development process. The robustness is evaluated with unseen data after the model
accuracy is estimated with cross-validation (CV), and regardless whether a Sony ERS-7 walks barefoot or wears socks, the forests achieve 94% accuracy. This result outperforms the state of the art techniques for indoor surfaces in the literature and the
classification execution is real-time on the robot. The above mentioned model development process with RF is documented to create new models for other robots more quickly and efficiently.
Kokoelmat
- Artikkelit [6140]