Apple generated HALO’s main training data using 250 NVIDIA H100 GPUs over approximately 48 hours, releasing training and deployment code.

The researchers reported synthesizing HALO’s principal training data using 250 NVIDIA H100 graphics processing units (GPUs) over approximately 48 hours. This figure applies specifically to the main data-generation stage, not the total duration or hardware requirements of the entire research pipeline.

Training used different hardware for distinct components: the motion generator was trained through imitation learning using 64 NVIDIA B200 GPUs. The motion tracker used four H100 GPUs for a process involving reinforcement learning and policy distillation; no corresponding training durations were supplied.

The released code covers data generation, model training, and deployment, including motion generation, motion tracking, task orchestration, and execution in simulation and on real robots. The sources report code availability but provide no exact repository URL, licensing terms, or statement that all generated datasets were released.

Simulation-generated data still required human input. The researchers said new physical tasks required manually written prompts and designed motion keyframes, while some motions also used human videos or motion-capture material.