WolfieWeb
Physical AI / Research frontier

Apple trained a robot in 1,000 virtual homes.
Then turned it loose in the real world.

HALO brings AI into unfamiliar rooms. The most revealing part is what happens when one successful movement has to become a sequence.

Real research footage · Click to enlarge
HALO research demonstration. Image credit: Apple / HALO research team. Original project.
1,000simulated training environments
360single-skill physical trials
91%reported single-skill success

A chair looks like an easy target. For a robot, it is a test of whether seeing, understanding, and moving can agree on the same physical world.

That is what makes this research worth watching. A convincing home robot must deal with ordinary furniture and awkward spaces. The interesting question is whether experience gathered in an artificial world can prepare a machine for a room outside it.

What Apple demonstrated

In a November 4 preprint, Apple researchers describe HALO on a Unitree G1. Training uses 1,000 generated environments; evaluation covers six unseen, unmapped real spaces. Skills include sitting, crawling, doorway passage, navigation, and box pushing. [1]

The bigger question

If useful physical experience can be generated at scale, how much of a robot’s education needs to happen in our homes?

Consider the potential advantage: a virtual practice room can be changed, reset, and tried again. A failed attempt does not have to become a dent in someone’s furniture. But the value of that practice depends on what survives the transition into reality. A beautiful simulation is useful only if the robot learns something it can actually use.

Watch the physical test

Look at the relationship between the robot and the furniture. Where is its body? Where is the obstacle? What must happen before the next move can begin?

Official Apple / HALO clips open in the WolfieWeb video lightbox. Videos stream from the research project. Full demonstration and original media ↗

From a request to a movement

HALO uses three layers: an orchestrator selects subtasks, a generator plans body motion, and a tracker executes it using depth and body-state feedback. [1]

A simplified view
01 / INTENTChoose a subtaskWhat should happen next?
02 / MOTIONPlan the bodyHow could the robot move?
03 / CONTROLFollow and adjustCan the body carry it out?

This division is a useful way to think about embodied AI. Understanding an instruction, proposing a motion, and executing it are different problems. The difficult part is making their answers remain compatible while the world keeps moving.

Official system diagram: Apple / HALO research team. Click to enlarge. [2]

The number behind the headline

Single-skill success averaged 91%. Three tested task sequences averaged 70%, with ten trials per sequence in one real environment. The hardest sequence reached 50%. These are different evaluations, not a controlled measure of a 21-point drop. [1, Table II]

Reported physical-world evaluation
EvaluationScopeSuccess
Individual skills360 trials / 6 spaces91%
Task sequences30 trials / 1 space70%

For a reader imagining a helpful household machine, the second test deserves attention. A useful errand is a chain of decisions. Finishing one movement creates the starting conditions for the next. Reliability has to last through the whole request.

Research, with clear limits

This is a preprint, not evidence of a shipping Apple household robot. The authors say new tasks still need human-authored prompts and keyframes; occlusion and contact remain weaknesses. [1]

Why WolfieWeb is watching

Our take: the most consequential milestone would be dependable adaptation. Can a robot encounter a different arrangement, recognize what matters, and complete the request without someone quietly repairing the plan?

HALO makes that question concrete enough to examine. Watch the footage, read the evaluation, and keep those two standards together. The next result worth getting excited about will be one that makes the entire sequence more dependable.

Study links, evidence & media credits

Research claims are attributed to the authors. Commentary is WolfieWeb’s interpretation. Reviewed November 8, 2026. WolfieWeb is independent of Apple and Unitree.

TOP ↑