← News

ResearchSources: arXiv · 16 Nov 2026

AGIBOT reports X-Nav humanoid navigation system completing indoor-to-outdoor missions

The X-Nav technical report describes a pipeline that takes a bipedal humanoid from a language instruction to a roughly hundred-metre walk from an office lobby to a store in another building.

In plain words

AGIBOT's X-Nav system lets a walking robot follow a plain-language instruction to travel from inside a building to an outdoor destination. Its MELIOR model steers the robot using only images from the robot's own camera, while an agentic workflow breaks the instruction into sub-tasks. The report says the robot completed a roughly hundred-metre mission on the Lingxi X2, including passing through a door and around obstacles.

What was announced

The report frames long-horizon indoor-to-outdoor navigation as an open problem for humanoids, since a single instruction must cover rooms, doors and roads while the camera moves with the robot's gait. X-Nav's data engine gathers navigation data in open-source scenes, in digital twins of deployment sites and on real robots. The MELIOR navigation model comes in two versions: MELIOR-A predicts actions and MELIOR-N predicts a navigation goal, both working from RGB images alone. The X-Navarch workflow converts a brief human instruction into sub-tasks and deploys MELIOR with explicit target navigation and obstacle avoidance.

On the Lingxi X2 robot, the system completed a mission of about a hundred metres from an office lobby to a store in a neighbouring building, with both MELIOR versions, and it followed a change of destination. LiDAR and odometry supported target approach and low-level control, and with X-Navarch the robot also passed static and dynamic obstacles. Before deployment, MELIOR was adapted with data from a digital twin of the site, which raised its closed-loop success rate in the twin; the authors also report competitive results on the R2R-CE and RxR-CE benchmarks.

What it does not show

  • The evidence is the paper's abstract only; the full report's methods, test conditions and success-rate figures are not available for verification.
  • Success-rate improvement from digital-twin adaptation is reported only within the digital twin, not measured independently at the real site.
  • The benchmark results are claimed by the authors and are not independently confirmed.

For security and IT teams

Scope
The described mission runs from an office lobby to a store in a neighbouring building on the Lingxi X2, with LiDAR and odometry used for target approach and low-level control.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.