Learning to Play Like a Pro: Adaptive Motion Planning and Tracking for Humanoid Robots
The decoupling trick is the part I keep coming back to. Most robot motion work treats planning and tracking as one fused problem, and the result is functional but stiff, like a robot that can get the ball in the goal but looks nothing like a person doing it. Splitting them so the plan carries the athlete's style and the tracker just keeps the robot on that plan under real-world noise is a smart way to get both grace and reliability. But the white paper is honest that this remains an open challenge, and I think the hard part is the speed-aware adaptive control.
A robot moving in slow motion can track almost anything. The moment you push it to game speed, the dynamics change, the joints saturate, the ground contact gets messy, and the beautiful planned trajectory falls apart. That's why the sim-to-real gap keeps eating humanoid robotics alive. I'd want to see how they're capturing the professional style in the first place, whether it's motion capture from real athletes or something synthetic, because that choice shapes everything downstream.
Where I land: this is the right problem to be working on. We've had humanoids that can walk and pick things up for a while. Making them move like someone who actually knows what they're doing, at speed, in a chaotic environment, is the next threshold. I'll be watching for the demo footage.