Research / Navigation

Can humanoid robots find their way around on their own?

Updated November 13, 2026 · 1 paper · How we rate evidence

Where things stand

Within a room, following directions in plain language, yes, in demonstrations. Long trips with several stops are still mostly measured in simulation: the strongest November 2026 result is 75.56% success on multi-goal trips in a simulator, with real-robot runs shown but not scored.

The newest approach is a vision-language-action model: the robot takes in camera images and a written instruction, such as walk to the table and then the door, and outputs walking commands directly.

Researchers are chaining short trips into long ones. Training on single-goal trips and then running multi-stop routes without extra training is a key test of whether these systems generalize.

Real-world results are thin. Few papers report success rates for long routes on a physical humanoid, and none we have tracked yet report navigation in a busy working site.

The papers

Moderate evidencePhysical robot and simulationUnitree G1

LHNav: chaining directions into long trips

LHNav: Long-Horizon Humanoid Navigation via Multirate Coordination and Physics-Grounded Action Data · Xueyang Wang et al. · arXiv 2611.12136 · Nov 9, 2026 · Preprint

What they did: A vision-language-action model that walks a humanoid to goals described in words, and chains several goals into one long trip.
Result: 75.56% success on long, multi-goal trips in simulation after training only on single-goal examples, and better than two earlier navigation models on single goals. On a real G1 it completed single-goal and multi-goal trips, including with rearranged objects.
Caveat: The headline number is from simulation; no success rates are given for the real-robot runs.

Checked against: Abstract. Results as reported by the authors; not peer-reviewed unless noted.

More research

All research Walking and balance Hands and handling objects Working with people Robot Rankings

How we pick and rate papers

We cover primary research that tests AI, perception, planning or control on a humanoid robot, or reports a distinct humanoid experiment. Product launches, funding news and promotional videos without a paper are left out.
Every paper links to its primary source. Most are arXiv preprints, which are not peer-reviewed; we say so, and add the journal or conference version when one is confirmed.
We check the numbers against the paper itself, and say whether we read the abstract or the paper text.
We separate physical-robot results from simulation, and note when a robot was teleoperated rather than acting on its own.
Strong means quantified tests on a physical humanoid with defined tasks and real comparisons. Moderate means limited hardware results or partial evidence. Preliminary means simulation-only or exploratory work.

Spotted an error, or a paper we should cover? Tell us, with a link to the paper.