Writing

Generative pretraining for motor control was shown at 680 hours of mocap on 24 A100s. A build log of the opposite experiment — under an hour of motion on one GPU, and the bugs that made RL silently not learn.

A reflection on three years across frontier robotics, biomanipulation, RL, FPGAs, and GPUs — and what it means to be nearing the natural close of that phase.

Reproducing a 3,553-hour bimanual behavior-cloning stack on a single GPU — which findings replicate at individual scale, which invert, and where the ceiling is.

Architecture, beamforming algorithms, and DSP techniques on hardware.

Why joint-order alignment mattered more than physics fidelity for my custom Tensaur robot in Isaac Lab.