HALO is a research system from Apple researchers that transfers movement skills trained in simulated homes to a physical Unitree G1 humanoid. The paper abstract describes training across 1,000 procedurally generated homes, without using a real home during training. The paper reports a 91% success rate across 360 trials in six unmapped real-world environments, while AppleInsider describes the same percentage as an average across real and simulated environments.
How HALO transfers skills from simulation
The researchers’ approach builds training examples from virtual homes, then applies the learned movement skills to a physical robot. The paper abstract says the training data was synthesized across 1,000 procedurally generated homes; AppleInsider reports around 300 hours of synthetic data and use of Nvidia Isaac Sim, a simulator for creating virtual environments.
AppleInsider identifies the physical platform as a Unitree G1 carrying a Mac mini with an M4 Pro chip and 64 GB of RAM. That hardware description belongs to this reported research setup, not an Apple-built humanoid product.
The paper’s skill list and the 91% result
The abstract names sitting, navigating to objects specified by name, stepping sideways through half-open doors, crouching and crawling under tables, and pushing boxes. It also describes combining skills into longer tasks. These are examples of HALO’s stated repertoire; the abstract does not tie every listed skill to the 360-trial result.
The headline success rate has two reported scopes. The paper abstract assigns 91% to 360 trials in six unmapped real-world environments. AppleInsider describes 91% as an average across real and simulated environments. The two accounts therefore do not describe the percentage on the same basis.
AppleInsider also reports that chaining tasks reduced success. That finding sits alongside the abstract’s description of longer composite tasks, but the two accounts do not provide a shared measure for comparing them.
What “autonomous” means here
A person still writes the prompt that tells HALO what task to perform. AppleInsider says no operator controlled the robot’s movements during the trials it described after that prompt was given. In this setup, autonomy applies to carrying out an assigned task, not choosing a goal independently.
For a robot, completing a task without real-time steering is different from deciding what task to do.
HALO is research using a Unitree robot, rather than an announcement of an Apple humanoid for consumers. The task-prompt requirement and the reported drop in success when tasks are chained mark the limits of the demonstrated setup.