Chunlok Lo and Greg Turk
Georgia Institute of Technology
Introduction
There have been numerous works in creating realistic character animations that is able to interact with the virtual world. These works mostly deal with abstracting the character into multiple articulated rigid bodies and driving its controls through joints using both more traditional PID controllers and recently, deep reinforcement learning techniques. However, while rigid bodies serve as a good approximation for vertebrates with skeletal structure, there have been little work done in animating characters without skeletons. Without skeletons, creatures must rely on volume preservation and muscle contractions for control. There have been several work dealing generating body plans and gaits for soft bodied creatures without skeletons[1][2][3], but they tackle the problem of generating a body plan rather than discovering a gait given a body plan.
In this work, we investigate the question of gait discovery for soft-bodied characters with specified body plan and muscle groups. We build on the soft-body simulator Voxelyze [4] to enable both isotropic and non-isotropic expansions of deformable voxels. We create a framework that is able to support multiple muscle groups and actuation patterns such as sinusoidal contractions. We then apply covariance matrix adaptation evolutionary strategy (CMA-ES) to find the best parameter of the contraction patterns to maximize our locomotion objective.
Methods
VoxCad
To simulate deformable creatures, we use a soft material simulator, Voxelyze[4:1], which is designed to be capable of modeling large deformations while being relatively stable and efficient. We modified the simulator to support non-isotropic expansions of voxels to better model of muscles which mostly contracts along a single axis. Check out my changes here. To better enable users to test various body plans with muscle contractions we use VoxCAD, a GUI editor designed for Voxelyze to create voxel shapes, and created additional mechanisms to create muscles, designated by the color of the voxel, and link muscles together into muscles groups to share parameters for their contraction patterns.
Designing a creature in VoxCAD. The colors correspond with muscles with the same material properties, contraction axis, and control parameters.
CMA-ES
Covariance matrix adaptation evolution strategy (CMA-ES) is a numerical optimization method that uses evolutionary strategies to sample according to a multivariate normal distribution. The main challenge with applying optimization strategy in this domain is the computation cost of the simulation. As a result, we chose to use CMA-ES over other algorithms like basic genetic algorithm or reinforcement learning to reduce the number of iterations required to obtain a result. Also given that we are able to determine the body plan and muscle groups prior to optimization, we can also limit the number of parameters for the creature to reduce computational cost of the optimization algorithm.
Contraction patterns
We experimented with sinusoidal contraction patterns as a starting point for this work. CMA-ES is able to optimize for the phase, amplitude, and frequency of the contraction to find the best configuration for locomotion. The user is also able to set the local axis by which the voxels within the muscle expands and contracts.
Results
Future Work
One avenue of future work is to experiment with other contraction patterns that better mirror real muscle acuations. Currently, the space of gaits possible are limited by sinusoidal actuations but better gaits might be possible. With more computational power, we can find a policy over a more complex parameter space that can perhaps allow for better gaits.
Another possible direction is to convert the current feedforward method into a feedback controller to allow the policy to respond to the environment. This could allow the user to control the direction the creature walks, and bypass obstacles that would otherwise be impossible.
Corucci, Francesco, et al. “Evolving Soft Locomotion in Aquatic and Terrestrial Environments: Effects of Material Properties and Environmental Transitions.” ArXiv:1711.06605 [Cs], Nov. 2017. arXiv.org, http://arxiv.org/abs/1711.06605. ↩︎
Nick Cheney, Robert MacCurdy, Jeff Clune, and Hod Lipson. 2013. Unshackling evolution: evolving soft robots with multiple materials and a powerful generative encoding. In Proceedings of the 15th annual conference on Genetic and evolutionary computation (GECCO '13), Christian Blum (Ed.). ACM, New York, NY, USA, 167-174. DOI: https://doi.org/10.1145/2463372.2463404↩︎
Evolved electrophysiological soft robots
N Cheney, J Clune, H Lipson ALIFE 14, 222-229 ↩︎
Hiller, J., & Lipson, H. (2014). Dynamic simulation of soft multimaterial 3d-printed objects. Soft Robotics, 1(1), 88-101. ↩︎↩︎
Soft Body Gait Discovery
Chunlok Lo and Greg Turk
Georgia Institute of Technology
Introduction
There have been numerous works in creating realistic character animations that is able to interact with the virtual world. These works mostly deal with abstracting the character into multiple articulated rigid bodies and driving its controls through joints using both more traditional PID controllers and recently, deep reinforcement learning techniques. However, while rigid bodies serve as a good approximation for vertebrates with skeletal structure, there have been little work done in animating characters without skeletons. Without skeletons, creatures must rely on volume preservation and muscle contractions for control. There have been several work dealing generating body plans and gaits for soft bodied creatures without skeletons[1][2][3], but they tackle the problem of generating a body plan rather than discovering a gait given a body plan.
In this work, we investigate the question of gait discovery for soft-bodied characters with specified body plan and muscle groups. We build on the soft-body simulator Voxelyze [4] to enable both isotropic and non-isotropic expansions of deformable voxels. We create a framework that is able to support multiple muscle groups and actuation patterns such as sinusoidal contractions. We then apply covariance matrix adaptation evolutionary strategy (CMA-ES) to find the best parameter of the contraction patterns to maximize our locomotion objective.
Methods
VoxCad
To simulate deformable creatures, we use a soft material simulator, Voxelyze[4:1], which is designed to be capable of modeling large deformations while being relatively stable and efficient. We modified the simulator to support non-isotropic expansions of voxels to better model of muscles which mostly contracts along a single axis. Check out my changes here. To better enable users to test various body plans with muscle contractions we use VoxCAD, a GUI editor designed for Voxelyze to create voxel shapes, and created additional mechanisms to create muscles, designated by the color of the voxel, and link muscles together into muscles groups to share parameters for their contraction patterns.
Designing a creature in VoxCAD. The colors correspond with muscles with the same material properties, contraction axis, and control parameters.
CMA-ES
Covariance matrix adaptation evolution strategy (CMA-ES) is a numerical optimization method that uses evolutionary strategies to sample according to a multivariate normal distribution. The main challenge with applying optimization strategy in this domain is the computation cost of the simulation. As a result, we chose to use CMA-ES over other algorithms like basic genetic algorithm or reinforcement learning to reduce the number of iterations required to obtain a result. Also given that we are able to determine the body plan and muscle groups prior to optimization, we can also limit the number of parameters for the creature to reduce computational cost of the optimization algorithm.
Contraction patterns
We experimented with sinusoidal contraction patterns as a starting point for this work. CMA-ES is able to optimize for the phase, amplitude, and frequency of the contraction to find the best configuration for locomotion. The user is also able to set the local axis by which the voxels within the muscle expands and contracts.
Results
Future Work
One avenue of future work is to experiment with other contraction patterns that better mirror real muscle acuations. Currently, the space of gaits possible are limited by sinusoidal actuations but better gaits might be possible. With more computational power, we can find a policy over a more complex parameter space that can perhaps allow for better gaits.
Another possible direction is to convert the current feedforward method into a feedback controller to allow the policy to respond to the environment. This could allow the user to control the direction the creature walks, and bypass obstacles that would otherwise be impossible.
Source Code [Github]
Corucci, Francesco, et al. “Evolving Soft Locomotion in Aquatic and Terrestrial Environments: Effects of Material Properties and Environmental Transitions.” ArXiv:1711.06605 [Cs], Nov. 2017. arXiv.org, http://arxiv.org/abs/1711.06605. ↩︎
Nick Cheney, Robert MacCurdy, Jeff Clune, and Hod Lipson. 2013. Unshackling evolution: evolving soft robots with multiple materials and a powerful generative encoding. In Proceedings of the 15th annual conference on Genetic and evolutionary computation (GECCO '13), Christian Blum (Ed.). ACM, New York, NY, USA, 167-174. DOI: https://doi.org/10.1145/2463372.2463404 ↩︎
Evolved electrophysiological soft robots
N Cheney, J Clune, H Lipson ALIFE 14, 222-229 ↩︎
Hiller, J., & Lipson, H. (2014). Dynamic simulation of soft multimaterial 3d-printed objects. Soft Robotics, 1(1), 88-101. ↩︎ ↩︎