cs.ROOct 10, 2026
SaveVbotNav-0: Asynchronous Cloud-Edge Dual-Policy Navigation across Indoor and Outdoor Environments
Abstract
Continuous indoor-outdoor navigation connects destination-level reasoning with motion decisions that must respond to current observations. We present VbotNav-0, a cloud-edge dual-policy framework that links these processes through asynchronous policy-to-policy conditioning. The cloud policy produces subgoal guidance from visual history, navigation instructions, and route tokens; the edge policy combines available guidance with current RGB and LiDAR observations to generate local trajectories. This formulation allows the local perception-action loop to continue between cloud updates. A common route representation connects indoor paths and outdoor map routes, and the training recipe combines navigation demonstrations, corrective data, and grounded visual question answering (VQA). In the reported closed-loop evaluation of 258 episodes across 145 scenes, VbotNav-0 achieves 55.04% success rate and 55.03% success weighted by path length.
Figures & tables
Figure 1: VbotNav-0 architecture and training. Upper left: System 2 aligns and tokenizes routes from navigation map software or spatial memory, and combines them with instructions and RGB history in a 4B cloud model to predict subgoals; the auxiliary VQA branch illustrates question–answer supervision. Lower left: System 1 fuses RGB semantics and geometry with LiDAR BEV features, compresses them with a Q-Former, and generates ten-point trajectories via a subgoal-conditioned flow-matching Transformer. The policies communicate asynchronously at nominal rates of 2 and 10 Hz. Right: Stage 1 combines mixed-source, multi-task SFT with DAgger; Stage 2 uses primarily real-robot data for trajectory-VAE learning and flow-matching-based local obstacle avoidance.
Figure 2: Schematic asynchronous schedule, initialized with usable guidance, not a latency measurement. A context snapshot is processed in the cloud and becomes usable only after arrival. Multiple edge decisions reuse the available guidance while incorporating fresh observations. A newer result updates subsequent decisions; an in-flight result cannot condition an earlier decision.
| 1 | Independently acquire cloud context at and generate guidance . |
|---|---|
| 2 | Make result eligible for edge use only after its arrival . |
| 3 | At each edge decision , acquire fresh observations . |
| 4 | If usable guidance exists, select the newest available source context and interpret it using . |
| 5 | Generate conditioned on and ; pass it to the locomotion controller. |
| 6 | Otherwise, defer to the system’s supervisory rule for missing or invalid guidance. |
Algorithm 1 Interface-level inference schedule corresponding to Eq. 2 . The schedule specifies dependencies, not a claim that timestamp ordering, guidance conversion, or fallback has been independently verified in deployment.
Figure 3: Reported navigation-data composition. Source and environment panels use nominal counts while retaining the supplied “+” labels. The scene shares are shown separately.
| Method | SR | SPL | Non-trav. | Timeout |
|---|---|---|---|---|
| CityWalker | 17.83 | 17.83 | 59.69 | 15.89 |
| ViNT | 15.12 | 15.12 | 58.53 | 0.39 |
| NaVid-VLN-CE | 8.53 | 8.53 | 20.54 | 69.77 |
| NaVILA | 9.30 | 9.30 | 25.58 | 62.79 |
| SocialNav | 17.83 | 17.82 | 70.16 | 1.55 |
| VbotNav-0 (Ours) | 55.04 | 55.03 | 38.76 | 3.49 |
Table 1: Closed-loop navigation results on 258 episodes across 145 self-collected scenes. All values are percentages; best results in each column are bold.
| Model variant | SR (%) | SPL (%) |
|---|---|---|
| VbotNav-0 (Ours) | 55.04 | 55.03 |
| w/o DAgger data | 38.03 | 37.98 |
| w/o VQA co-training | 45.81 | 45.68 |
Table 2: Ablation of DAgger data and VQA co-training. Each ablated variant removes the indicated component from VbotNav-0. Both metrics are percentages; higher is better. Best results are bold.
| Model variant | Road rules Pass rate | Traversable classes F1 | Region localization [email protected] | Waypoint diagnosis Pass rate |
|---|---|---|---|---|
| Base (w/o VQA) | 33.0 | 40.5 | 12.4 | 3.5 |
| VQA-trained | 59.3 | 87.2 | 59.1 | 45.0 |
Table 3: Effect of VQA training on semantic understanding and spatial grounding. All values are percentages ( ); best results are bold.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Spatial grounding before and after VQA training. Original boxes and IoU annotations are preserved. The crosswalk example retains an extra prediction.
Figure 5: Waypoint diagnosis before and after VQA training. The blue reference routes are shown only for evaluation.
Figure 6: Road-rule understanding before and after VQA training. Answers are condensed from the supplied examples; original observations are preserved.
Figure 7: First predictions on a bridge approach. VbotNav-0 selects a traversable goal along the bridge corridor. CityWalker, SocialNav, and ViNT select non-traversable endpoints; NaVid and NaVILA initially produce accepted turn-derived goals but later stall.
Figure 8: First predictions on a plaza route. VbotNav-0 selects a traversable goal along the route and succeeds. CityWalker and SocialNav select non-traversable first endpoints. The other baselines have accepted first endpoints but fail later in their closed-loop rollouts.
Figure 9: First predictions on a stair approach. VbotNav-0 selects a goal within the annotated traversable corridor and succeeds. SocialNav and ViNT select non-traversable first endpoints. CityWalker initially selects an accepted endpoint but fails later, while NaVid and NaVILA subsequently stall.
Figure 10: Selected snapshots from an indoor–outdoor navigation rollout. From top to bottom, the robot proceeds through a clear doorway (frame 350), initiates a left turn when the maneuver becomes locally actionable (frame 1202), follows a later right turn at an open intersection (frame 7480), and stops after reaching the route endpoint (frame 7766). In each row, the left panel overlays the predicted local trajectory on the egocentric view and reports the grounded rationale and action; the right panel shows global route progress, current pose, start, and goal.
More in this topic
Robot Navigation
Robot navigation studies how robots perceive, localize, plan, and move safely toward goals in static or changing environments.
See all 532 papers