Newborn animals in the wild must learn to walk quickly to avoid predators . Aside from the innate reflexes of the spinal cord to help the animal avoid falling during their initial attempts, they tend to learn relatively quickly. However, the details of how this happens are still not well understood . So why not build a four-legged robot to better understand the details?
A robot to understand how animates learn to walk
This is exactly what the researchers of the International Max Planck Research School for Intelligent Systems (IMPRS-IS) did. To explore how animals learn to walk, they built a four-legged robot the size of a dog as part of a research study geared towards understanding its details. The result was published in the journal Nature Machine Intelligence with the title “Learning plastic matching of robot dynamics in closed-loop central pattern generators” .
Called Morti, the robot consists of four bi-articular legs mounted on a carbon fiber body. Each leg has three segments: femur, stem and foot. The segments of the femur and foot are connected by a spring-loaded knee joint to mimic the biarticular muscle-tendon structure.
For animals, learning to precisely coordinate muscles and tendons takes time : initially puppies rely heavily on more basic reflexes, which help the animal during the first attempts at movement, while the subsequent muscle control, the more advanced one, needs some refinement.
Under the hood of Morti, we find sensors mounted on the feet to measure contact with the ground and several position and speed sensors. Leading it all is a Bayesian optimization algorithm that guides learning using foot sensor information coupled with target data from a virtually modeled spinal cord, which runs as a software application on a computer.
Various optimizations were implemented sequentially on Dead, with performance assessed by continually comparing sent and predicted sensor information and adapting engine control patterns in response. The robot started making good use of the leg mechanism after a duration of about an hour:
“Our robot is practically 'born' without knowing anything about the anatomy of the legs or how they work (…) the CPG [central pattern generator] looks like a built-in automatic walking intelligence that nature provides and that we have transferred to the robot. The computer produces signals that control the leg motors and the robot initially walks and trips. The data flows back from the sensors to the virtual spinal cord where the sensor and CPG data are compared (…) If the sensor data does not match the expected data, the learning algorithm modifies the walking behavior until the robot walks well and without stumbling. Changing the CPG output while keeping the reflexes active and monitoring the robot stumbling is a fundamental part of the learning process "
In the future, the researchers intend to extend the CPG by taking body pitch into account when generating trajectories. With a unit of inertial measurement, the pitch of the body could be inserted into the CPG: in the current formulation, the CPG does not assume the inclination of the body and relies on the robustness that passive elasticity adds to the system to compensate for the inclination of the body. existing body.
Furthermore, Morti bases its operation on inexpensive hardware (<€ 4,000) with low computing power (5 W, that required by a Raspberry Pi ) and with control frequencies of 500 Hz (control loop) and sensors with lower frequencies. compared to the state of the art of other locomotion controllers based on more complex models and requiring high bandwidth computation and high control frequencies.