The new Atlas robots are all electric VS the old series that were electro-hydraulic. As a "retirement" thing, Boston Dynamics released a video compilation of Atlas' robots worst fails, including many burst hydraulic hoses that they'd never shown before. Apparently a large part of designing all the stunts was managing the weight of the robots VS the pressure spikes landing would cause in the hydraulic systems. The motions would have to be designed to spread the impact force out over enough time to avoid popping something.
On Wednesday, July 16, 2025 at 11:41:17 AM MDT, Chris Albertson <albertson.ch...@gmail.com> wrote: Actualy for this they use motion capture from a human dancer to collect “target points.” The points are (x,y,z, time). Then, of course, the robot does not have the same mass distribution as a human and would fall down if it exactly copied the human’s moves. So they use an “MPC” controller. MPC is like the far simpler PID, but MPC looks to minimise predicted future error, not the current error. OK, not the “future error” but the “future cost”. Cost is the difference between where you are and where you want to be. An MPC controller can parallel park a car given only the current position (in a traffic lane) and desired location (at the curb). It searches for the lowest cost solution and will find the back-and-forth motions and steering wheel turns needed. The robot searches in real time for the best solution for hitting those target points, even if it means widely deviating from them. The cost funtion is coded by engineers and likely has a low cost for slightly missing a point and VERY high cost for falling down or for exceeding the motor’s limitations. The real-time optimizer finds a path that has the lowest predicted integrated cost from present to some future time, usually a few seconds ahead. “integrated cost function” is a very mathematical-sounding term, but simply said, it is placing its foot now, so as to create the lowest cost future. Basically the controller “thinks” “I was told to place my foot at (xy,z) at time “t", but if I do that I will fall on my face, so I place it differently, but then again, I need to minimize the effort needed to reach the next foot postion and so on. MPC is using a physics-based simulation to predict about 10,000 possible futures and then selecting the one that is “best”. Then it does this all over again about 50 times a second. The controll pretty much does what humans and other animals do. We think ahead a little. We will walk around a tree, not through it, because the longer path creates a better future. The robot can in theory be told “build a brick wall” and then search for the motions that create the desired state. But not today. computers are not powerfull enough This new Atlas robot is nothing more than about 28 BLDC servo motors. I'm sure many of us could design the mechanical system. Where Boston Dynamics shines is in its mastery of advanced control theory. They say it is based on this: https://en.wikipedia.org/wiki/Model_predictive_control _______________________________________________ Emc-users mailing list Emc-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/emc-users
