Learn to walk for a machine dog for an hour

Use learning algorithm to optimize virtual spinal cord

Source ｜ Science and Technology Daily

Author ｜ Zhang Mengran

Edit ｜ Zhang Runqi

According to the "Nature · Machine Intelligence" magazine on the 18th, in order to understand how animals learn to walk and learn from trip, researchers in Max Planck Intelligent Systems (MPI-IS) in Germany built a four-foot machine dog " "Modi", it learned to walk in just one hour.

Moti made full use of complex leg mechanics to learn through the guidance of the Bayesian optimization algorithm: its foot sensor information matches the target data of the modeling virtual spinal cord running in the machine dog. Machine dogs learn to walk by constantly sending and expected sensor information, running reflexes, and adjusting their motor control mode.

Among humans and animals, the central mode generator (CPG) is a neuron network in the spinal cord, which can produce periodic muscle contraction without brain input. It helps to generate rhythmic tasks, such as walking, blinking or digestion. Machine Dog Modi optimizes its movement mode faster than animals in about an hour.

During the steady walking of the machine dog, the sensor data from its feet continued to compare with the expected landing of the machine dog CPG. If the machine dog trip is tripped, the learning algorithm will change the distance between the leg back and forth, the speed of the leg swing, and the length of the leg on the ground. The adjustment of movement will also affect the ability of the machine dog to use its leg mechanics.

During the learning process, the CPG sends a adjustable motor signal so that the machine dog can never trip and optimize its walking.

The first author of the thesis, the former doctoral student of the MPI-IS Dynamic Movement Research Group, said: "Our machine dog is actually" born ', CPG is similar to the built-in automatic walking intelligence provided by nature, We have transferred it to the machine dog. When the data flows back to the virtual spinal cord from the sensor, compares with the CPG data. If the sensor data does not match the expected data, the learning algorithm will change the walking method until the machine dog walks well and it will not trip stumbling. fall."

Chief Editor Circle Point

The central mode generator is a biological nerve ring that produces animal rhythm movement. It is a complex distributed neural network composed of neuroshicas and multiple reflection circuit systems. In recent years, CPG -based robotics control has become a new research hotspot in the field of bionic robots. Researchers let the four -foot robots learn to walk in a very short time. To understand it, they learn while walking. Its CPG is constantly simulating, and the learning algorithm is constantly adjusted according to the differences between the data from the sensor and the simulation data. Artificially planned robot gait is too stiff, while this biological gait allows four -foot robots to better adapt to the surrounding environment.

contact number:

(010) 82081166-6059

Breaking the mailbox:

[email protected]

China Net Technology

Pay attention to the various hotspots of the science and technology circle, covering character dynamics, corporate stories, black technology products, scientific knowledge ... This is the gathering place for technology enthusiasts.

- END -