← Paper Digest

X-Nav Technical Report: Real-World Long-Horizon Indoor-to-Outdoor Navigation on a Humanoid Robot

Jiaxin Huang, Chengxing Lin, Xiaoyu Luo, Sirui Chen, Jingyu Guo, Aolin Li, Kairui Chen, Rui Zhen, Cheng Wen, Kerui Gu, Shunyao Wang, Zhihui Peng, Shaoli Huang

November 2026

TL;DR AgiBot's X-Nav takes a Lingxi X2 humanoid on a language-instructed 96.6 m walk from an office lobby to a store in the next building, with no map and only the built-in RGB camera driving its vision-language navigation model. Adapting the model in a Gaussian-splat twin of a site raises closed-loop success within 3 m from 40% to 80%.

What they do

Vision-language navigation models score well on 10-15 m indoor benchmark episodes, but real humanoid runs have stayed short. X-Nav takes a bipedal Lingxi X2 about a hundred metres from an office lobby, through a door and along a road to a store. It works from a natural-language instruction, with no map or reference route, and its navigation model sees only the built-in head camera.

How it works

  • Data engine. About 6.3M navigation samples replayed in an Isaac Sim platform from R2R, RxR, SRDF and 1,095 open-source scenes. Digital twins are scanned deployment sites rendered with Gaussian splatting, where the policy’s own rollouts are labelled with recovery and stop corrections. Recordings on the real robot capture how the gait moves the camera.
  • MELIOR. A vision-language model fed the instruction, the current frame and up to seven past frames. MELIOR-A outputs up to four discrete moves; MELIOR-N outputs a pixel or polar goal, a turn, or a stop. Training is supervised fine-tuning, then DAgger corrections in simulation.
  • X-Navarch. A multimodal planner splits the instruction into sub-tasks and reviews each stop. MELIOR-X finds and approaches named objects with an open-vocabulary detector, a verifier, SAM3 tracking and LiDAR range. A goal adapter uses odometry to keep goals fixed despite inference delay, and an avoidance layer handles obstacles.

Results

  • With MELIOR-A, the robot covered 96.6 m (34.2 m indoors, 62.4 m outdoors) in 3 min 24 s. It followed a destination change at 9.1 m and waited 24 s for a car and pedestrians. A second run with X-Navarch and MELIOR-N passed three dynamic and two static obstacles.
  • Site adaptation in the twin of a lobby (100 episodes): adding correction windows raised success within 3 m from 40% to 80%. Episodes ending in sustained collision fell from 52 to 3. Success within 1 m did not improve (28% to 20%).
  • R2R-CE val-unseen: MELIOR-N reaches 70.5% success and 65.1% SPL, the best among single forward-RGB methods. On RxR-CE it reaches 60.9%, below the strongest methods.

Limitations

The real-robot evidence is one route on one floor, reported as two runs rather than success rates. Stairs, lifts and doors that must be opened are out of scope. Site adaptation is tested at one site, with the same scene and target in training and evaluation, and precise stopping is still open. Inference runs on an off-board GPU.

Figures

From Figure 1: Top, a mission of about a hundred metres on the Lingxi X2, from an office lobby to a store in a neighb...
From Figure 1: Top, a mission of about a hundred metres on the Lingxi X2, from an office lobby to a store in a neighbouring building. Bottom, the three parts of the system: the data engine (open-source scenes, digital twins, real robot), the MELIOR navigation models, and the X-Navarch workflow with obstacle avoidance.
From Figure 11: The X-Navarch workflow. A planner splits the instruction into sub-tasks; MELIOR-N follows a route or ...
From Figure 11: The X-Navarch workflow. A planner splits the instruction into sub-tasks; MELIOR-N follows a route or MELIOR-X approaches a named target; a goal adapter and an obstacle-avoidance layer pass each goal to the locomotion controller.
From Figure 12: The mission with MELIOR-A. The robot gets a new destination after 9.1 m, crosses the lobby, leaves th...
From Figure 12: The mission with MELIOR-A. The robot gets a new destination after 9.1 m, crosses the lobby, leaves through a glass door, waits while a car and pedestrians cross, and stops at the store entrance after 96.6 m.
From Figure 6: Digital-twin collection in an office lobby: the Gaussian-splat twin, 500 planned tasks, the 48,006 tra...
From Figure 6: Digital-twin collection in an office lobby: the Gaussian-splat twin, 500 planned tasks, the 48,006 training windows, and closed-loop results with and without correction windows.
Abstract

Long-horizon navigation from indoors to outdoors remains an open problem for humanoid robots: one instruction covers rooms, doors and roads, and the camera moves with the gait. Worken, AGIBOT's measure of productive robot work, depends on robots that reach their workplaces autonomously. We present X-Nav, to our knowledge the first system to take a bipedal humanoid through a language-instructed long-horizon mission from indoors to outdoors. X-Nav covers a pipeline from data to deployment. Its data engine collects navigation data in open-source scenes, in digital twins of deployment sites and on real robots. Its navigation model, MELIOR, takes only RGB images from the robot's built-in camera and comes in two versions: MELIOR-A predicts actions, and MELIOR-N predicts a navigation goal. Its agentic workflow, X-Navarch, turns a brief human instruction into sub-tasks and deploys MELIOR on the robot with explicit target navigation and obstacle avoidance. On the Lingxi X2, X-Nav completes a mission of about a hundred metres from an office lobby to a store in a neighbouring building; LiDAR and odometry are used for target approach and low-level control. The mission is completed with both versions of MELIOR: the robot follows a change of destination and walks through a door from the lobby to the road, and with X-Navarch it also passes static and dynamic obstacles on the route. During deployment, MELIOR is adapted with data collected in a digital twin of the site; the adaptation raises its closed-loop success rate in the twin, and the adapted model is then deployed on the robot at the real site. In addition, MELIOR achieves competitive results on the R2R-CE and RxR-CE benchmarks.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.