Distributed-learning control for soft peristaltic robots
A peristaltic robot advances inside a tube the way an earthworm does: a wave of expansion travels along its body, anchoring and pushing it section by section. In the group's design that wave is obtained from a single air input, so control has two variables available, fill time and vent time. This project proposes a controller that learns which combination of times produces the desired speed in the duct the robot is in, and that extends to several robots sharing information. It worked in simulation. It could not be validated on the physical setup because the robots failed before the tests were completed.
Context
The robot is a silicone part with five cavities connected by a channel and fed from one end. When air is injected, the cavities inflate in sequence, and when it is released they deflate. Each cycle leaves a net displacement. The concept comes from two earlier projects in the lab: the design and modeling of the robot and its manufacturing process.
Longitudinal section of a prototype: five cavities connected by a channel
Displacement per cycle depends on the diameter and friction of the duct, which are not known in advance and may change along the way. A fixed pair of times that works in one tube may not work in another.
Controller
The controller is formulated with imitation dynamics, a tool from evolutionary game theory. An abstract population is distributed among nine strategies, corresponding to increasing, keeping or decreasing each of the two times by a fixed step. Each strategy receives a fitness that is highest when the resulting speed matches the reference. The population migrates toward the fittest strategies and the action applied to the robot follows from that distribution.
Because the robot's analytical model did not predict its actual behavior well, fitness is not computed from the model. The robot tries each strategy and measures its progress with a camera. Learning is slower, but it does not depend on a model that is not trusted.
Architecture for four robots: one controller per robot, Arduino-driven valves and cameras to measure progress
In the multi-agent case, each robot has its own controller and the robots share information, so that the one with the smallest error influences the others.
Simulation results
In Simulink, with the duct radius changing from 1 to 1.5 cm halfway through the simulation, the robot's speed approaches the reference and recovers after the change.
Robot displacement in simulation with the controller active
Evolution of the share of the population in each strategy
Convergence is slow: after about 50 cycles some error remains. With three robots starting from different conditions, all three converge toward the same reference.
Physical tests
Several robots were cast in silicone with paraffin inserts, and the system was set up for two robots in parallel, with valves, Arduino and cameras.
Paraffin insert inside one of the molds
Physical setup for two robots in parallel
The tests revealed three problems the model did not anticipate:
- At a certain inlet pressure, the inflation order reverses and the robot moves backward.
- With a single air source, pressure is shared among the open valves.
- Thin walls puncture. Four repair methods were tried; the most effective was paper impregnated with silicone, which seals but stiffens the repaired area. No commercial adhesive worked.
No experiment was achieved that combined intact robots with the controller in operation.
What is missing
The controller is specified and tested in simulation, but its validation was left pending because of a manufacturing problem. The author proposes shortening the robot to reduce buckling of the insert, using a stiffer core, and defining a reliable repair method.
How it fits in Robiolab
The project belongs to the lab's first soft-robot line and shows a difficulty that recurs in it: the soft body does part of the control work by itself, since the inflation sequence is encoded in the geometry, but in exchange the behavior is hard to model and very sensitive to fabrication. The controller responds to that by learning from trial rather than relying on the model, an approach that remains relevant for the group's current continuum robots.