Mowgli

2004–2008

Introduction

An animal’s musculoskeletal system gives it the ability to actively move about in a wide variety of environments. Jumping and landing movements are characterized by large instantaneous forces and short duration. Here we present an approach to realize motor control of jumping and landing which exploits the synergy between control and mechanical structure. Our experimental system is a bipedal robot with an artificial musculoskeletal system including mono-articular and bi-articular muscles. A McKibben type pneumatic muscle is used for the actuator.
jumping
Fig.1 Mowgli jumping

Mowgli (original)

Overview

Mowgli weighs about 3kg, is 0.9m body height with the legs extended, and has 6 pneumatic muscles for 6DoF legs. Based on biomechanics of biological musculoskeletal structure, the robot has the tapered legs. In other words, the proximal (close to trunk) muscles are bigger than distal muscles.
Mowgli original
Fig.2 Mowgli: muscle layout and appearance

Jumping and Landing

Mowgli can jump as high as 0.5m, which is more than 50% of its body height. This is extremely high for legged robot with multiple-DoF. Although motor commands are given simultaneously, we observed a proximo-distal sequence of joint extensions. The results show that the passive dynamics of the body are the dominant factor in dynamic motion.
jumping onto the chair
Fig.3 Jumping onto a chair with a height of 0.4m

Mowgli2

Overview

Mowgli 2 is a improved version of the original Mowgli robot developed for motor learning. The improved points of the Mowgli 2 are use of proportional control valves and additional muscles on the hip joint. The robot weighs about 3 kg, is 0.84 m body height with the legs extended, and has 8 pneumatic muscles and 4 passive springs for legs. A leg has 3 DoF, one for each joint (hip, knee, and ankle). Several proportional pressure control valves and a CPU board are mounted on the robot. For the purposes of the learning experiment, which requires several hours, electrical power and compressed air is supplied from external equipment. The robot has an orientation sensor, a potentiometer on each joint, a pressure sensor on each muscle, and a touch switch on one foot.
Mowgli 2
Fig.4 Mowgli2: muscle layout and appearance

Computer Simulation

We perform simulation experiments by using a simplified model of the robot in the pilot study. There is a modelling error between the simulation model and the real robot. In addition, the real robot has high uncertainties concerning impact force regarding landing. Therefore, we proposed computationally-efficient motor learning method with small numbers of trials for the real robot.
running snapshots
Fig.5 Pilot study: simulation experiments.

Learning jumping motion

We employ the mixture of random exploration and improved hill-climbing search as the optimization technique for the learning. The system uses the hillclimbing search start with the pattern obtained from previous random exploration. We apply the method to a planar jumping with 4 muscle groups. The left leg and right leg receive same activation. The parameters to be learned consist of continuous values of air pressure and duration time for each phase. The number of phases of activation pattern is set to 2. We conducted three learning sessions which includes 100 or 150 trials including 50 random exploration trials. The evaluation function is a linear combination of maximum height and the similarity between the observed position at the end of the movement to the desired squatting position. The activation patterns are valued for each jumping trials. The vertical jumping was acquired in all of these 3 sessions. Learning time for one search session was about 30 minutes.
running snapshots
Fig.6 Snapshots from the random exploration.
running snapshots
Fig.7 Acquired movement: jumping and soft landing

References

Contribution

Ryuma Niiyama organized the project, under the supervision of Professor Yasuo Kuniyoshi. Ryuma Niiyama designed all hardware, electric circuits, and controllers. The simulation study of the robot was conducted by Kei Kakitani and Ryuma Niiyama.

Papers

Ryuma Niiyama, Kei Kakitani and Yasuo Kuniyoshi
Learning to Jump with a Musculoskeletal Robot using a Sparse Coding of Activation
In Proceedings of the ICRA 2009 Workshop on Approaches to Sensorimotor Learning on Humanoid Robots, pp.30–31, Kobe, Japan, May 2009.
PDF
BibTeX

@INPROCEEDINGS{Niiyama2009_Learning-to-Jump-with-MusculoskeletalRobot,
author = {Ryuma Niiyama and Kei Kakitani and Yasuo Kuniyoshi},
title = {Learning to Jump with a Musculoskeletal Robot using a Sparse Coding of Activation},
booktitle = {Proc. ICRA 2009 Workshop on Approaches to Sensorimotor Learning on Humanoid Robots},
year = {2009},
pages = {30--31},
address = {Kobe, Japan},
month = {May},
}

Ryuma Niiyama, Akihiko Nagakubo and Yasuo Kuniyoshi
Mowgli: A Bipedal Jumping and Landing Robot with an Artificial Musculoskeletal System
In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA 2007), pp.2546–2551 (ThC5.2), Roma, Italy, April 2007.
PDF
BibTeX

@INPROCEEDINGS{Niiyama2007_Mowgli-BipedalJumping-and-LandingRobot,
author = {Ryuma Niiyama and Akihiko Nagakubo and Yasuo Kuniyoshi},
title = {Mowgli: A Bipedal Jumping and Landing Robot with an Artificial Musculoskeletal System},
booktitle = {Proc. IEEE Int. Conf. on Robotics and Automation ({ICRA 2007})},
year = {2007},
pages = {2546--2551 ({ThC5.2})},
address = {Roma, Italy},
month = {April},
}