Organizations: AgiBot X-Lab (Boson Lab) · MBZUAI · UESTC · University of Macau · Nanyang Technological University · The University of Melbourne · Wuhan University
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.
Figures & tables
System
Output
Input
Site prior
Highlight
Qwen-RobotNav ( 2026c )
waypoints
RGB
none
15.6M training samples
Robostral Navigate ( 2026 )
image waypoints
single RGB
none
reinforcement learning
NavGPT-3 ( 2026a )
tool calls, actions
RGB, map
none
reasoning planner above an action policy
VbotNav-0 ( 2026 )
sub-goal, trajectory
RGB, LiDAR
route or memory
cloud–edge policy, indoor and outdoor
HoloAgent-0 ( 2026b )
skill calls
RGB or LiDAR
spatial memory
persistent 3D spatial memory
X-Nav (ours)
actions or navigation goal
single RGB
none
∼ 100 m indoor-to-outdoor humanoid mission
Table 1: X-Nav and recent navigation systems.
Figure 2: The three loops of X-Nav . The sub-task loop (L1), the mission loop (L2) and the deployment loop (L3). The box below each loop gives one result of this report: the success rate (SR) of MELIOR-N on R2R-CE, the length of the mission on the Lingxi X2, and the share of closed-loop episodes in a digital twin that stop within 3 m of the goal without and with correction windows.
Name
Part
Description
X-Nav-Sim
data engine
Simulation platform for data collection and evaluation.
X-Nav-Syn
data engine
Navigation data collected in open-source scenes.
X-Nav-Twin
data engine
Navigation data collected in the digital twin of a deployment site.
X-Nav-Real
data engine
Onboard RGB and robot motion recorded on the Lingxi X2.
MELIOR-A
navigation model
Predicts up to four actions per decision.
MELIOR-N
navigation model
Predicts one navigation goal per decision.
Table 2: Components of X-Nav .
Figure 3: The three routes to navigation data. Collection in open-source scenes covers diverse layouts and tasks. Collection in the digital twin supplies data for site adaptation. Collection on the Lingxi X2 captures the robot’s onboard observations and physical motion. Each tier is labelled with its dataset, X-Nav-Syn , X-Nav-Twin , or X-Nav-Real ( Section 3.3 ); the base tier gives the samples of X-Nav-Syn and the total of open-source scene collection, which includes the replayed routes of R2R, RxR, and SRDF.
Figure 4: The generation pipeline of X-Nav-Syn . (1) Goal-pose sampling. We select target instances and sample goal poses that pass reachability, visibility, and clearance checks. (2) Offline planning. We sample start-goal pairs within route quotas, plan paths with A* on the occupancy grid, convert them to action sequences, and check the resulting routes. (3) Replay & annotation. We replay the reference routes to record RGB-D observations and camera poses. A vision–language model generates instructions. We derive action labels from the routes and navigation goal labels from the replayed poses and depth. We combine the observations, instructions, and labels into step samples.
Figure 5: Trajectory collection in X-Nav-Sim . Top: the collection steps, shared by open-source scenes and digital twins ( Figure 4 gives the details). Left and right: one collected example of each task type, goal-directed navigation and route following, with the occupancy grid of the scene, the reference route in the region of the dashed box, the first-person views at the three numbered positions, and the generated instruction. Middle: goal poses of one target, a dining table. With one goal pose, routes from different starts end at the same pose and share the final view. With the goal poses that pass the pose checks, labelled audited goal poses (G1–G4), routes from different directions end at different poses, each with the target in view. Triangles mark starts, circles indicate goal poses and headings, and the star marks the target.
Figure 6: Digital-twin collection in an office lobby. (1) The digital twin of the scanned lobby, rendered with 3D Gaussian splatting. (2) 500 tasks planned on the occupancy grid of the twin; the lines are reference trajectories. (3) The 48,006 training windows of the 500-trajectory setting: windows from reference trajectories, recovery windows, and stop-boundary windows. (4) Closed-loop evaluation in the twin for an adapter trained on reference trajectories only and for one trained with correction windows ( Section 7.2 ).
Figure 7: Supervised fine-tuning data for MELIOR-N . (a) Samples per dataset on a radial axis (rings at 1M, 2M, and 3M samples), with the four navigation datasets at the top, the two visual–language understanding datasets at the bottom, and the total in the centre. (b) Training mix: navigation samples account for 73.8% and visual–language understanding samples for 26.2% of the 8.55M samples. (c) Sample counts by dataset.
Figure 8: The two versions of MELIOR. Both read the instruction and current and history observations of one forward RGB camera. At each decision step, MELIOR-A predicts up to four primitive actions, and MELIOR-N predicts a single navigation goal in one of four modes ( Section 4.3 ). In the action sequence, F is a 0.25 m forward step, L and R are 15∘ left and right turns, respectively, and S stops the robot. Bottom: decisions of the two versions on the Lingxi X2, seen by the forward camera. For MELIOR-A , the marker shows where the predicted actions lead. For MELIOR-N , the circle marks the pixel goal, and the curve from the lower edge to the goal is drawn for display.
Figure 9: The navigation-goal interface of MELIOR-N . Top: the four modes of the navigation command and their tokens. visible marks a pixel (u,v) of the current view with range r , bearing ϕ and final heading ψ ; polar gives a goal outside the view; rotate turns in place; stop declares the instruction complete. Middle: from one frame to a walking robot. The model predicts a navigation goal, the LiDAR ranges its pixel, the goal is anchored in odometry, the obstacle-avoidance layer takes the robot around static and dynamic obstacles, and the controller walks; the next frame is taken after about four seconds ( Sections 6 and 6 ). Bottom: pixel goals (circles) of MELIOR-N on the forward camera of the campus route; the curve from the lower edge to the goal is drawn for display.
Figure 10: Overview of MELIOR-X . Given a language query, the skill (a) grounds and verifies the target, (b) tracks its mask, and (c) estimates its distance d and bearing θ from the LiDAR points inside the mask to approach it; (d) near the target it switches to the downward RGB-D view. The tag of a stage gives its update rate. The images are renderings of the lobby twin with illustrative overlays.
Figure 11: X-Navarch , the agentic workflow of X-Nav . The planner divides the mission instruction into sub-tasks and reviews their outcome at hand-off. One navigation skill is active at a time: MELIOR-N follows a route, and MELIOR-X approaches a named target. On the robot, the goal adapter converts a navigation goal into a goal pose, and the obstacle-avoidance layer passes it to the locomotion controller.
Figure 12: The mission with MELIOR-A . The Lingxi X2 starts in an office lobby (1) and receives a new destination after 9.1 m, a convenience store in a neighbouring building (2). It crosses the lobby (3), leaves through a glass door (4), turns right along the road (5), waits while a car and pedestrians cross (6), and stops at the store entrance after 96.6 m (8). The bar gives the positions of the frames on the route; distances are cumulative odometry.
Figure 13: The mission with X-Navarch and MELIOR-N , with obstacles on the route. X-Navarch receives the mission instruction, replaces the destination when a new instruction arrives (between 1 and 2), and drives the Lingxi X2 with the navigation goals of MELIOR-N . The robot passes a cart that is pushed across the lobby (2) and a sweeping robot (3), walks through the side door (4), passes a robot parked at a kiosk (5), yields to a pedestrian (6), and passes through the gap between a sign and traffic cones (7) before it reaches the store (8). The note on the right of a frame gives the type of obstacle.
Table 3: Closed-loop results in the digital twin of the lobby (100 episodes). SR@ r : episodes that end within r of the goal; STOP-SR@ r : episodes that also end with a predicted stop; SPL and STOP-SPL weight the two by path length. NE in metres, all other columns in per cent. Both rows are adapters of MELIOR-A .
Figure 14: The Lingxi X2 with the adapted model in the real lobby. The robot crosses the open floor, passes the lockers and reaches the vending machines; the number on the right of a frame is the time since the start.
R2R-CE
RxR-CE
Method
NE ↓
OSR ↑
SR ↑
SPL ↑
NE ↓
SR ↑
SPL ↑
nDTW ↑
ONE FORWARD RGB STREAM
NaVid ( 2024 )
5.47
49.1
37.4
35.9
–
–
–
–
Uni-NaVid ( 2025a )
5.58
53.3
47.0
42.7
6.24
48.7
40.9
–
NaVILA ( 2025 )
5.22
62.5
54.0
49.0
6.77
49.3
44.0
58.8
StreamVLN ( 2026b )
4.98
64.2
56.9
51.9
6.22
52.9
46.0
61.9
Table 4: R2R-CE and RxR-CE, val-unseen. NE: navigation error in metres; OSR, SR, SPL and nDTW in per cent. Results of other methods are taken from the cited papers; a dash marks a value that is not reported. Bold marks the best value of the upper group; the MELIOR-A row is the mean of two checkpoints. The lower group lists methods with panoramic input, a planner with map tools, or training at a much larger scale.
Figure 15: R2R-CE and RxR-CE val-unseen. MELIOR-A and MELIOR-N compared with three methods of the upper group of Table 4 ; in each group the bars follow the order of the legend.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Group
Tokens
Meaning and encoding
Control
5
<visible> , <polar> , <rotate> , <stop> , and <end>
Pixel
256
Uniformly spaced coordinate values on the 640×480 frame, shared by u and v
Range
32
Bins uniformly spaced in log(1+r) over 0–15 m; larger distances map to the final bin
Angle
72
Circular values at 5∘ intervals, shared by ϕ and ψ
Total
365
Appendix
Table 5: Navigation token vocabulary of MELIOR-N .
Following a route (implicit navigation)
Approaching a target (explicit navigation)
Sub-task
a route described in language
a named object to approach
Component
MELIOR-A , MELIOR-N
MELIOR-X
Target representation
implicit, in the weights and the context of the model
explicit state: mask Mt and base-frame position ptB
Input
instruction; RGB only: the current frame and at most seven earlier frames
query; RGB with LiDAR, or RGB-D at close range
Language model
runs at every decision
runs when the target is acquired or reacquired
Output
up to four actions, or one navigation goal
distance and bearing, then a velocity request
Appendix
Table 6: The two types of sub-task and the navigation skills that execute them.
Table 7: X-Navarch on the Lingxi X2: planner and execution of navigation goals.
Item
Setting
Locator
LocateAnything-3B; up to three candidate boxes per image and query
Verifier
compact vision–language model ( Section 7 ); score from the first-token logits of the two answers ( Equation 3 )
Acceptance
score of at least 0.5; highest score chosen; boxes that cover more than 60% of the image rejected
Initialisation
two consecutive acceptances for the same query; the second box initialises the tracker
Tracker
SAM3 mask propagation on ordinary frames
Reacquisition
changed query, explicit reset, or three consecutive frames without a valid mask
Appendix
Table 8: MELIOR-X : acquisition, tracking and ranging settings.
Figure 16: Design of the learned local policy. The policy operates between the navigation goal and the frozen locomotion controller; the colour of an arrow is the stage that adds it. Stage 1: a navigation policy with one backbone for all robots maps the local goal, the encoded surroundings and the body state to commands. Stage 2: a body encoder for each type of robot. Stage 3: a reactive policy with a safety gate for moving obstacles. The executed motion returns as body state (dashed).