How 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°.

We tried a whole zoo of ways to learn five-ball juggling throw corrections. The simplest thing on the board came out on top.

The progression. Three balls on the first attempt, four and five by the second.

Learning 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