Robot dog learns to walk in one hour

A four-legged robot that is just learning to walk wanders around like a young animal. However, the robot learns to walk fluidly in about one hour, whereas a foal or giraffe would require much more time. The artificial representation of the animal's spinal cord is a computer program, which quickly learns to optimize the robot's movement. The artificial neural network quickly self-adjusts and is not initially perfectly tuned.

To evade predators, a baby giraffe or foal must learn to move as quickly as possible on its legs. Animals have networks for coordinating their muscles in their spinal cords from birth. However, it takes some time to master the perfect synchronization of the tendons and muscles of the legs. Animal young first rely significantly on hardwired reflexes in the spinal cord. The animal's motor control reflexes, however considerably more primitive, enable it avoid falling and harming itself when it first tries to walk. After then, it is necessary to exercise more complex and exact muscle control until the nervous system has finally become fully attuned to the young animal's leg muscles and tendons. The juvenile animal is no longer uncontrollably stumbling; instead, it can now keep up with the adults.

To better understand how animals learn to walk and learn from mistakes, scientists at the Max Planck Institute for Intelligent Systems (MPI-IS) in Stuttgart undertook a study. They created a four-legged, canine-sized robot to aid in their analysis of the situation.

According to Felix Ruppert, a former doctorate student in the Dynamic Locomotion research group at MPI-IS, "As engineers and roboticists, we sought the answer by constructing a robot that displays reflexes much like an animal and learns from mistakes." "Is it a mistake if an animal stumbles? Not if it just occurs once. However, if it trips up frequently, it tells us how well the robot walks."

"Learning Plastic Matching of Robot Dynamics in Closed-loop Central Pattern Generators," which will be released on July 18, 2022 in the journal Nature Machine Intelligence, has Felix Ruppert as its first author.

Virtual spinal cord optimization using learning algorithm

Ruppert's robot makes effective use of its intricate leg mechanics after learning to walk in just one hour. The learning is guided by a Bayesian optimization algorithm, which compares the target data from the modeled virtual spinal cord running as a program in the robot's computer with the measured foot sensor information. By executing reflex loops, comparing sent and expected sensor data, and modifying its motor control patterns, the robot gradually learns to walk.

A Central Pattern Generator's control settings are adjusted by the learning algorithm (CPG). These central pattern generators in both humans and animals are networks of neurons in the spinal cord that cause regular muscle contractions without brain input. Networks with central pattern generators help create rhythmic actions like blinking, walking, or digesting. Furthermore, brain connections that are hard-wired and connect sensors in the leg to the spinal cord cause reflexes, which are involuntary motor control activities.

The movement impulses from the spinal cord can be controlled by CPGs if the young animal walks on a completely level surface. However, a slight unevenness in the terrain alters the gait. In order to prevent the animal from falling, reflexes take over and modify the animal's gait. These brief shifts in the movement signals are reversible, or "elastic," and after the disturbance, the movement patterns resume their pre-disturbance state. But if, despite active responses, the animal continues to stumble throughout several cycles of movement, new movement patterns must be learnt and become "plastic," or irreversible. When an animal is a baby, its CPGs are initially not tuned properly, causing it to stumble on both smooth and uneven ground. But the animal quickly picks up on how its CPGs and reflexes manage the muscles and tendons in its legs.

The same is true for "Morti," a Labrador-sized robot dog. Furthermore, the robot can optimize its movement patterns in approximately an hour, which is quicker than an animal can. A portable, lightweight computer that manages the movement of the robot's legs simulates Morti's CPG. The location of this simulated spinal cord is where the head would be on the back of the quadruped robot. Sensor data from the robot's feet is continually compared with the anticipated touch-down predicted by the robot's CPG during the hour it takes for the robot to walk smoothly. The learning system modifies the length of time a leg is on the ground, the speed at which it swings, and how far it swings back and forth if the robot trips. The robot's ability to make effective use of its flexible leg mechanics is also impacted by the modified motion. The CPG transmits modified motor impulses to the robot during the learning phase to reduce stumbling and improve walking. In this architecture, the motors, springs, and leg design of the robot are not explicitly known to the virtual spinal cord. It is missing a robot "model" and has no understanding of the mechanics of the device.

Our robot essentially "births" with no knowledge of its limb anatomy or function, according to Ruppert. "The CPG is similar to a naturally occurring autonomous walking intelligence that we have transmitted to the robot. The robot first moves and falls because the computer sends out signals that control the motors in its legs. The virtual spinal cord receives data from the sensors and compares it to CPG data. The learning system modifies the robot's walking behavior if the sensor data does not match the predicted data until the robot can walk steadily and without tripping. A crucial step in learning is altering the CPG output while maintaining reflexes and watching for robot stumbles."

Power-saving robot dog control

Only five watts are used by Morti's computer when it is moving. Industrial quadruped robots from well-known manufacturers are substantially more power-hungry since they have become adept at moving with the aid of sophisticated controls. Using a model of the robot, their controls are programmed with knowledge of the robot's precise mass and body geometry. They generally use a few tens of watts to several hundred watts. Both robot kinds function dynamically and effectively, but the Stuttgart model uses far less processing power. Additionally, it offers crucial insights on animal anatomy.

"A living animal's spinal cord is difficult to study. However, we can simulate one in the robot "explains Alexander Badri-Spröwitz, who leads the Dynamic Locomotion Research Group and co-authored the article with Ruppert. "These CPGs are recognized in a wide variety of mammals. We are aware that reflexes are ingrained, but how can we combine the two to enable animals to learn actions through both CPGs and reflexes? This is crucial investigation into the interface between biology and robotics. The robotic paradigm provides us with solutions to issues that biology alone is unable to address."

Story Source:

Materials provided by Max Planck Institute for Intelligent Systems. Note: Content may be edited for style and length.