Kai Ploeger @kaiploeger.bsky.social · 07/07/2026A prior on how to adapt is much more forgiving than a prior on what the optimal policy is. A simplified Newtonian model may struggle to predict exactly how to throw a ball, but it is very useful for deciding how to adjust the aim. 000
Kai Ploeger @kaiploeger.bsky.social · 07/07/2026How accurate does a model-based prior over search directions have to be? We rotated the analytic Jacobian prior to test this. At 90°, it fails as expected. But surprisingly, up to 30° the learning speed is barely affected, and final performance stays unchanged even up to 45°. 100
Kai Ploeger @kaiploeger.bsky.social · 02/07/2026 Paper and videos: kai-ploeger.com/residual-jugglingkai-ploeger.comTask-Error Residual Learning for Real-Robot Five-Ball JugglingStable three-, four-, and five-ball juggling from the second attempt, via directional task-error residual learning. 000
Kai Ploeger @kaiploeger.bsky.social · 02/07/2026The only method that came close is a far heavier composite Bayesian-optimization scheme, and it shares exactly those two ingredients. Same formulation, much more machinery. 100
Kai Ploeger @kaiploeger.bsky.social · 02/07/2026The pattern underneath: Keep the direction of the error instead of collapsing it to a scalar, and guide exploration with an analytic prior pointing roughly in the correct direction. 100
Kai Ploeger @kaiploeger.bsky.social · 02/07/2026We tried a whole zoo of ways to learn five-ball juggling throw corrections. The simplest thing on the board came out on top. 120
Kai Ploeger @kaiploeger.bsky.social · 30/06/2026Break this rule on purpose, enter the beautiful world of siteswap juggling, as my former student Mario Gomez Andreu did. His bachelor's thesis on it became an IROS paper, with mesmerizing videos: kai-ploeger.com/beyond-casca... 000
Kai Ploeger @kaiploeger.bsky.social · 30/06/2026Fun fact: Patterns with odd and even ball counts are juggled differently. Odd counts cross from hand to hand (the cascade), even counts stay on their own side (the fountain). This is the simplest way to keep two balls out of the same place at the same time. 100
Kai Ploeger @kaiploeger.bsky.social · 30/06/2026The progression. Three balls on the first attempt, four and five by the second. 121
Kai Ploeger @kaiploeger.bsky.social · 24/06/2026It runs on a deliberately simple stack: soft PD control, idealized ballistics, no visual perception in the control loop. A task-level residual absorbs the rest. 010
Kai Ploeger @kaiploeger.bsky.social · 24/06/2026Learning five-ball juggling on the second attempt, with two Barrett WAMs. Most humans take years of practice. Paper and videos: kai-ploeger.com/residual-juggling #robotics #juggling 145