Editorial: Soccerbots take a dive
DELIBERATELY taking a tumble to implicate another player might earn you a yellow card on the soccer pitch, but falling over safely is an essential skill for all soccer players, both human and humanoid.
Advertisement
Now soccerbots in Chile are learning to fall in a controlled way, reducing damage to themselves and their environment, letting them recover quickly and get on with the game- they can even fall over deliberately to save a shot on goal, for example (see gallery).
of the University of Chile in Santiago and his team are training two teams of soccerbots for the annual world . The aim of the RoboCup project is to have a robotic team in place by 2050 that can beat the best human side 鈥� an impossible goal unless soccerbots can learn to fall over without damaging themselves.
Before this, most research on humanoid robotics has concentrated on keeping the robots upright and balanced. But in the real world they are certain to lose their balance and fall over, or trip up on uneven surfaces, says , professor of artificial intelligence and robotics at the University of Sheffield in the UK.
When the inevitable happens, most robots fall as a dead weight, unaware that anything has gone awry until their sensors tell them that they are horizontal. And unlike humans, they don鈥檛 try to break their fall or protect themselves from the impact.
Inspired by the controlled falling taught in some martial arts, Ruiz-del-Solar wanted to find the best ways for a robot to keel over. He and his team have come up with a set of equations that quantify how much total damage a fall is likely to cause, given the speed with which the robot crashes to the ground, the forces and torques transmitted to each of its joints and the position of its cameras and other important body parts relative to the ground. So a good fall will be one that generates a low overall damage number from their equations.
To find out the optimum ways for a robot to fall, the team used a computer simulation based on a humanoid robot called Nao, the player used by all teams competing in the RoboCup鈥檚 Standard Platform League. Nao has 22 simple joints, each with a single degree of freedom, and is typical of the bipedal soccer robots being built today.
Ruiz-del-Solar and colleagues put their simulated soccerbot through a series of different fall sequences. The simulation computes the stresses on each joint, which can then be plugged into the team鈥檚 equations to work out the total damage factor.
They found that one of the main ways to minimise damage is for the robot to fold its legs underneath it. Among other things, that means the robot is much less likely to hit its head on the ground. Another good strategy is to use a fall sequence consisting of several movements, so the falling body has several points of contact with the ground, spreading the energy of the impact over a large number of joints, rather than taking it all in one disastrous crunch.
鈥淥ne of the main ways to minimise damage is for the robot to fold its legs underneath it鈥�
The Santiago team tested their method for real using their UCH H1 robot (pictured above), which they built to compete in the RoboCup鈥檚 Humanoid League. It is similar to Nao but has a stronger frame and joints. Using a high-speed camera, they recorded the speed and acceleration of the robot鈥檚 joints as it fell, and used that to calculate the forces and torques on each joint. The tests confirm that UCH H1 suffers less damage when it bends its legs to keep its centre of mass low as it falls. The work is to be published in Robotics and Autonomous Systems ().
While the fine details of the best tumbling strategies will vary from robot to robot, and depend on how the tumble starts, Ruiz-del-Solar is encouraged to find that the results for real robots mirror what happened in the simulation 鈥� despite the robots being different models. It suggests that the equations should be general enough to apply to any humanoid robot with standard joints, he says.
Ruiz-del-Solar estimates that a well-equipped soccerbot would need about five different fall sequences stored in its memory, to be triggered when the robot is fouled or needs to dive to save a shot on goal, for example.
The true test of the new work will come at this year鈥檚 RoboCup, to be held in Graz, Austria, in June and July, where the Santiago team are planning to try out their robot, programmed with safe fall sequences, on the soccer pitch.
After that, they hope to develop an algorithm that would calculate from first principles the best way for the robot to fall, removing the 鈥渢rial and error鈥� element of their approach.
Sharkey describes the work as 鈥渘ovel and promising鈥�. He points out that giving a large robot the ability to fall over gracefully could greatly extend the role of humanoid robots outside soccer, including and that assist walking in elderly and disabled people. 鈥淎t present if someone tripped in a cybersuit they would fall heavily and sustain injuries,鈥� he says. Improving the technology could make a fall less harmful for anyone wearing a such a suit. 鈥淲hat we need now is a good method for getting up gracefully.鈥�
Editorial: Soccerbots take a dive
