Physix FrontierPhysix Frontier · News Briefing Card (ifanr · Nov 9, 2026)
Apple unveils HALO research: Unitree G1 with Mac mini runs autonomous home tasks
KEY FACTS
- Apple's research team has published HALO, a humanoid robot study that uses a Unitree G1 body carrying a Mac mini M4 Pro.
- The system completed 360 trials across 6 unscanned real homes, with an average success rate of 91%.
- All training data was synthesized in simulation, comprising 163,000 samples and nearly 300 hours.
- Generating 50,000 sit-down skill samples consumed 250 H100 GPUs running for 9 hours.
- The orchestrator runs the Alibaba Qwen3.6-35B-A3B model locally by default.
KEY DATA
91%Average success rate in real environments
163,000Simulation-synthesized sample count
Nearly 300 hoursSynthetic data duration
About 70%Composite task success rate
PHYSIX OBSERVATION
Apple replaces human data collection with simulation synthesis, handing the data bottleneck to compute - and that is more noteworthy than the robot body itself. The mix of a Unitree body, an Alibaba model, Nvidia chips and a Mac mini shows that humanoid robot competition is shifting from hardware dexterity to who can accumulate experience more cheaply. But with composite task success at only 70%, it is still far from home practicality.
Source: ifanr report