Tactile Feet and Online Learning for Robust Control of a Quadruped Robot in Challenging Terrain

点击翻译按钮获取中文摘要

Tactile sensing has been extensively explored in the domain of object manipulation and detection, especially with the TacTip tactile sensor, but very little has been done in the domain of tactile feet for walking robots. With feet there is a similar issue as in hands: by interacting with the environment the robot occludes the surface that needs quantifying (whether this be in texture, surface orientation, edge detection etc). Therefore the application of a tactile sensor, which directly measures the contacted surface, seems more appropriate than commonly used non-contact methods such as computer vision (CV), lidar or sonar which cannot sense the surface under contact. The addition of tactile feet may give a walking robot more reliable estimation of slip between the feet and floor, floor distance and orientation and identification of smooth/bumpy surfaces (for selection of stable footholds), all of which improve the success of a quadruped walking and/or running over uneven natural terrain. Most walking robots struggle to do this. Tactile feet pose additional challenges to those faced in tactile hands. The hardware must be capable of supporting the body weight and withstand the impact from walking/running and be resistant to unexpected objects encountered in natural terrain (e.g. sharp objects). The software must handle the greater distortion and noise of the sensor caused by the movements of the supported robot, which is inherently more complex and dynamic than the noise experienced in hands, and the feedback loop must be faster as walking or running require faster movements than with object grasping and manipulation. Additionally given the size and mobile nature of a walking robot, any software must either run on an onboard microcomputer, which restricts computational power, or there must be a high speed connection to an offboard computer which is fast enough to enable real time control of the robot. Additionally, as explored in the first year project, the use of online learning methods to learn the mapping between tactile data and useful control inputs could be a data efficient and robust way of training the robot. Offline methods tend to be extremely data intensive, needing data across all possible dimensions of variation, and cannot learn to adapt to novel inputs encountered during testing. Additionally the models are learnt for specific sensors which may break and need replacing at any time, and while replacement sensors will be similar there is enough difference that the same models cannot be used and all training must be repeated. The use of online learning solves these problems, introducing instead the challenge of creating comprehensive online data collection policies in order for the correct mappings to be learnt autonomously. A neat extension from the first year project, which explored data-efficient online sensor training and control in edge following tasks, is the development of a beam-balancing robot. Fitting a quadruped with tactile feet would enable it to walk along a randomly curving beam only slightly wider than its leg span, without falling. With tactile sensing it is possible to identify an edge and how far along the sensor the edge is (as explored with a single sensor in the first year project). This is in contrast to a robot without tactile sensing which would be unable to tell if the foot is fully on the beam or not - the use of onboard cameras watching the feet would be obscured from seeing directly under the feet by the feet or the terrain (e.g. grass) so may be unreliable. Using tactile sensors in this way would allow the robot to move the feet back onto the beam and prevent falling (in a more complex system using maps the locations of the edge could be noted to aid future planning of stable courses).