Apple's HALO humanoid trains in 1000 simulated homes, hits 91% success in real houses
heyshrutimishra · x · 2026-11-10
Apple's HALO system pairs a Unitree G1 humanoid with Alibaba's Qwen models running locally on a Mac mini strapped to the robot — no cloud, no internet. Training used zero teleoperation data: 1000 simulated homes, then zero-shot transfer to 6 unseen real homes, achieving 91% success across 360 trials (sitting on sofas, pushing boxes, following spoken commands). The data economics are the unlock: 250 GPUs produced 300 hours of training data in 48 hours, replacing hour-per-hour human capture. Chaining 3 tasks still drops to 50% success. Unitree stock jumped 8%, Alibaba up to 4% on the paper.
More from Embodied
- At CoRL in Austin? Take a Tesla robotaxi ride, says robotics researcher — pduan · 2026-11-10
- Machenta demos instant robot learning at CoRL 2026, founded by Imperial College one-shot learning team — Ed__Johns · 2026-11-10
- NVIDIA Unveils GR00T 2 Open-World Action Model, Two Papers Shortlisted for CoRL Best Paper — danfei_xu · 2026-11-10
- Nomagic Team Headed to CoRL 2026 to Dig Into Physical AI Mastery — m_wulfmeier · 2026-11-10
- ROBIB Meetup #3 in Ghent to Debate Build-vs-Buy for Humanoid Robots — MarwaEldiwiny · 2026-11-10
- CoRL 2026 kicks off in Austin, Texas, running November 9-12 — shaohua0116 · 2026-11-10