Apple unveils HALO: fully autonomous humanoid with VLM orchestration, 84% chair-sit success
carlosdponx · x · 2026-11-06
Apple's AIML research team published HALO, a fully autonomous perceptive humanoid system trained in simulation and deployed on real hardware, with paper, video and code.
The autonomy stack has three layers:
- Orchestrator (0.5 Hz, backpack compute): an off-the-shelf VLM turns spoken requests into subtasks (push box, follow me, sit inside, crouch under table), advancing on success and retrying on failure.
- Generator (3 Hz): from RGB image and current subtask, a frozen Qwen3.5-2B encoder with a DiT decoder and flow matching (15 Euler steps) produces 5 s of whole-body motion (150 frames, 36 values) with real-time chunking.
- Tracker (50 Hz, embedded compute): follows the plan using onboard depth and proprioception to output targets for 29 joints.
Demos include consecutive spoken commands and tasks like sitting on a chair/sofa, crawling under desks, sidestepping through doors, and following a person — sit-task success rate is 84%.
Related event: Apple Unveils HALO Humanoid Robot Research With 84% Sitting Success Rate(2 posts)→
More from Embodied
- GPT-6 Astra tops a 16-benchmark robot index, yet went 0/11 on real plug insertions — AIFlow_ML · 2026-11-06
- Giving Claude a cutting-edge robot body just so it can spoon-feed me — repligate · 2026-11-06
- moonlake previews research on scaling robot learning with simulation in the RL era — du_yilun · 2026-11-06
- Nearly 100 testers praise Sphenomenon MR demo that turns a whole house into voxels — Scobleizer · 2026-11-06
- Spectral Skills lands on real robots: composing motion skills via energy functions — daibond_alpha · 2026-11-06
- Developer gives Claude Desktop a body and sensors via MCP and Raspberry Pi — jfrankie · 2026-11-06