Progress, newest first.

RoboDuck's walking policy, trained with reinforcement learning in MuJoCo. At a 0.1 m/s command with a 0.12 floor it tracked 101.5% ± 2.1 of the target across 3 seeds × 10 runs with zero falls. Simulation only; no hardware yet.

We pooled four public two-arm towel-folding datasets and trained the same small policy on the pooled data and on one dataset alone. Pooled 10-step action error: 0.1442. Single dataset: 0.1321. Hold-position baseline: 0.1744. Next we test per-dataset normalization.

These are offline metrics on recorded data, not robot success rates. No hardware yet.