Sensory Integration and Decision Making Based on Insect Brain Model

This work sought to advance understanding of how to construct and control tensegrity spines: highly compliant robots with many degrees of freedom inspired by biological spines. Our initial goal was to control a spine or snake-like tensegrity robot with central pattern generators (CPGs), based on the intuition that these two oscillatory systems would synchronize. Our intuition suggested the compliance of its tensegrity structure should give the robot a high degree of locomotion competence at the structural level, which would complement the adaptability of the central pattern generator based controller. Since analytical design methods for tensegrity robots were not (and still are not) well defined, we decided to test the idea in a preliminary version of the NASA Tensegrity Robotics Toolkit (NTRT). I started with simple sine waves providing a trajectory to distributed impedance controllers for the morphology that became Tetraspine, and eventually moved to an open loop CPG. In simulation, movement was achieved with suspicious alacrity on rough terrain, including hills, blocks, a ramp, and a wall as discussed in Tietz et al. 2013. Due to the success in simulation we decided to pursue hardware in parallel, with one version of Tetraspine at CWRU and another version at NASA Ames. Ultimately the hardware was capable of locomotion, but not of the terrain handling we observed in simulation due to the spine's shorter length. With results from our first hardware attempt, we were able to make improvements to NTRT including replacing the default position-based cables from the Bullet physics engine with a custom Hooke's law based spring model, resulting in more realistic though less impressive simulations. Based on Tetraspine's performance under these new conditions, we concluded it would be worth investigating additional morphologies. Tom Flemons, a tensegrity researcher from British Columbia, has developed a series of passive tensegrity models inspired by vertebrate anatomy [Flemons2007, Flemons2012]. With four different morphologies, it would have been too time consuming to hand tune controllers for each experimental morphology. Thus, I developed an algorithm for automatically determining a CPG controller based on the morphology of a tensegrity structure, assuming reasonable symmetries between the actuators. We then adapted machine learning techniques such as two step Monte Carlo to parametrize these controllers, including code for parallel processing of trials. I applied this technique to four morphologies controlling between forty four and eighty eight actuators with up to seventy two degrees of freedom, a record high for a tensegrity robot, as discussed in Mirletz et al. 2014. In addition to contributing to NTRT's open source release, I then added contact dynamics to our custom cable models. To my knowledge, this is currently the only open source cable model with both realistic forces and contact dynamics. This enabled reinvestigation of rough terrain locomotion with improved realism. As expected, the success of the tensegrity spines depended on the ratio between the size of the obstacle and the size of the spine. In order to enable spines to handle larger obstacles, I adapted feedback equations for a central pattern generator based controller in a novel way [Mirletz, Quinn, and SunSpiral 2015]. I added descending inputs via an artificial neural network for goal direction, a step towards allowing for simple `go that way' commands from an operator [Mirletz et al. 2015 b]. In order to verify the realism of these results, Dr. In-Won Park and I used the NASA Ames Tetraspine hardware to verify the forces predicted by NTRT. We showed that NTRT is capable of predicting the maximum tension in the hardware within 7.9%, giving us confidence in our ability to use NTRT to design future robots [Mirletz et al. 2015 a]. To summarize, this work has developed the first CPG based robotic control capable of goal directed locomotion on rough terrain, the first controller for a tensegrity robot capable of goal directed locomotion on multiple types of terrain, and a validated simulator for tensegrity robots. This work will hopefully inform and encourage future work in compliant, adaptive robotic systems with many degrees of freedom, particularly robots with flexible spines and torsos. Future robots with flexible spines and torsos will be better able to perform in situations such as search and rescue, planetary exploration, and environments dangerous to humans.