Part of the Robotic Movement series. Project page: Humanoid movement through physics and optimization.
Technique beats size
I played tennis in college, coached in college, and directed tournaments before that. The thing that stayed with me from all of it is the serve. You can have a huge, strong, athletic player whose serve will not break 50 mph, and a short, skinny kid with years of training who hits 100. Strength is not the bottleneck. Technique is.
The trained kid serves harder because he uses a kinetic chain: from his feet to his calves, his quads, his hips, his shoulder, and only then his arm and hand. The untrained player uses his arm. No matter how much you weigh, if you cannot move that weight well, it is mechanically useless. The kid who uses his whole body keeps accelerating the racquet head through the whole movement, so in a real sense he has more mass behind the shot than the bigger player. The comparison led me to focus on coordinated motion and the transfer of energy through the body.
Nobody has perfect technique. The best players are simply the ones closest to the ideal. And the ideal has a shape: use as many muscle groups as you can, in a sequence synchronised well enough that you never lose acceleration.
From a serve to a theory of movement
The serve was only ever the example. Lifting a weight, sprinting, jumping, throwing: every repeatable movement that can be mechanically optimised follows the same principle. What I wanted was not a model that gets better at one movement after a million repetitions, but something closer to a generalized theory of movement mechanics, where in every situation the body looks for its most efficient way to produce the result, whether that is hitting a serve or washing dishes.
Why a robot, not a player
For a person, predicting an optimal movement is difficult: segment masses, joint ranges, muscle properties, and motor control are only partially measured. Coaches use experience and observation to bridge those gaps.
For a humanoid robot, every one of those numbers is already written down in its specification. The inverse problem nobody can solve for a human is tractable for a robot. That is the move this project makes: give a robot its own specification and a physical goal, and let it find its own technique.
There is a second reason. Most humanoids today learn by copying people: humans in motion-capture suits, with the motion retargeted onto the robot. But a human optimum is optimal for a body of tendons and muscle, not for electric motors with completely different torque-speed characteristics and no elastic energy storage. That difference motivates a comparison between retargeted human motion and strategies derived from the robot’s own body.
From the serve to the throw
The project started with the vertical jump, the most studied maximal human movement. When it came time for a second task, I chose a throw. It is the closest analogue to the serve I could simulate without a racquet: the same independent muscle groups integrated into one kinetic chain, without the engineering of a racquet-ball impact. The throw turned out to be the moment the whole project came into focus.
Where it went
The project has grown well past what I first expected. It started as a research question; it could become a company. A presentation or a workshop is one way to show it, but the work is not built around any single venue. It is built around one question: can a body work out how to move from a description of itself?
Next in the series: Jumping.