Apple’s HALO trains a Unitree G1 on 300 hours of synthetic motion and reports 91% real-world success

Eight Apple researchers, led by Siyuan Hu with Vladlen Koltun as final author, posted “Perceptive Humanoid Motion from Simulation” to arXiv on November 4, 2026. Their system, HALO, runs on a stock Unitree G1 fitted with a wide-angle camera and a Mac mini on its back, untethered. Trained mostly on motion generated in simulation, it walks to objects, sits on chairs, steps through doors, crouches and crawls under tables and pushes boxes in rooms it has not seen, with a 91% average success rate across 360 real-world trials.

Why it matters: it is a public humanoid-robotics result from Apple with code released, and it adds to the evidence that whole-body behaviour can be learned largely from generated data rather than teleoperation.

How is HALO built?

Three layers run at different speeds. A vision-language model splits requests into subtasks about every two seconds. A diffusion model of about 49 million parameters plans five seconds of whole-body motion about three times a second. A reinforcement-learning tracker running at 50 Hz on the G1’s onboard Jetson Orin NX turns that plan into targets for 29 joints. Training motions came from a motion-diffusion model placed in rooms generated with Infinigen (45% living rooms, 30% offices, 25% bedrooms) and rendered in NVIDIA Isaac Sim. Some human data was still used, including motion-capture clips.

What did the ablations show?

Training on one room instead of 1,000 cut simulated sitting success to 43% from 84%, and replacing the diffusion decoder with plain regression cut it to 19%, according to Humanoids Daily’s reading of the paper. The authors note that the skill set grows only as fast as people can specify new tasks. Background on the approach: our sim-to-real explainer and DexForce’s simulated data.