Date of Completion
ecological psychology, postural coordination, haptics, self-organization, artificial intelligence, recurrent neural networks, tensegrity
Field of Study
Doctor of Philosophy
Tensegrity systems have been proposed as both the medium of haptic perception and the functional architecture of motor coordination in animals. However, a full working model integrating those two aspects with some form of neural implementation is still lacking. A basic two-dimensional cross-tensegrity plant is designed and its mechanics simulated. The plant is coupled to a Recurrent Neural Network (RNN). The model’s task is to maintain postural balance against gravity despite the intrinsically unstable configuration of the plant. The RNN takes only proprioceptive input about the springs’ lengths and rate of length change and outputs minimum lengths for each spring which modulates their interaction with the plant’s inertial kinetics. Four artificial agents are evolved to coordinate the patterns of spring contractions in order to maintain dynamic equilibrium. A first study assesses quiet standing performance and reveals coordinative patterns between the tensegrity rods akin to humans’ strategy of anti-phase hip-ankle relative phase. The agents show a mixture of periodic and aperiodic trajectories of their Center of Mass. Moreover, the agents seem to tune to the anticipatory “time-to-balance” quantity in order to maintain their movements within a region of reversibility. A second study perturbs the systems with mechanical platform shifts and sensorimotor degradation. The agents’ response to the mechanical perturbation is robust. Dimensionality analysis of the RNNs’ unit activations reveals a pattern of degree of freedom recruitment after perturbation. In the degradation sub-study, different levels of noise are added to the RNN inputs and different levels of weakening gain are applied to the forces generated by the springs to mimic haptic degradation and muscular weakening in elderly humans. As expected, the systems perform less well, falling earlier than without the insults. However, the same systems re-evolved again under the degraded conditions see significant functional recovery. Overall, the dissertation supports the plausibility of RNN cum tensegrity models of haptics-guided postural coordination in humans.
Fultot, Martin, "Tensegrity and Recurrent Neural Networks: Towards an Ecological Model of Postural Coordination" (2020). Doctoral Dissertations. 2616.