VbotNav-0: Asynchronous Cloud-Edge Dual-Policy Navigation across Indoor and Outdoor Environments

• The design targets coordination between route-level reasoning over visual history and instructions and local motion responsive to obstacles, traversability, and pose.Source: §1 · p. 1
• Cloud policy guidance conditions edge trajectory generation alongside fresh onboard observations, while the edge recomputes trajectories between cloud results.A separate locomotion controller executes the edge output.Source: §1 · p. 1
• The framework uses a common route interface for indoor planner paths and outdoor map routes, with lookahead tokens describing upcoming segments.Source: §1 · p. 2
• Its stated focus is a common route-conditioned interface between navigation policies and explicit accounting of guidance available at each edge decision.This separates the communication schedule from local trajectory updates and supports evaluation of cloud-update delays.Source: §1 · p. 2
• 55.04% SR and 55.03% SPL are reported across 258 episodes and 145 scenes.Source: §1 · p. 2
• The contributions include asynchronous conditioning that distinguishes context acquisition, result arrival, and local decisions, plus empirical study of navigation, training components, and grounded understanding.Source: §1 · p. 2

• Prior systems explore shared representations, spatial interfaces, high-level commands, cached or delayed guidance, and asynchronous cloud–edge conditioning.The cited works make asynchronous high-level guidance an established design direction; VbotNav-0 examines it with a common route interface and two navigation policies.Source: §2 · p. 2
• Navigation work covers language-guided indoor tasks, semantic exploration, heterogeneous-data goal reaching, outdoor street instructions, route guidance, social reasoning, and environment transitions.VbotNav-0 uses a supplied reference route, instructions, and visual observations; its distinction is the information available to each policy, not pretraining-corpus geography.Source: §2 · p. 2
• Corrective data and grounded VQA complement navigation imitation, with VbotNav-0 evaluating navigation and understanding separately rather than treating answer quality as a direct physical-safety measure.Source: §2 · p. 2

• The cloud receives an instruction, visual history including current and historical frames, and route tokens; the edge receives current RGB and LiDAR observations.Source: §3 · p. 3
• The route window contains lookahead geometry in the robot frame and is derived from the supplied route, not future observations or executed ground-truth motion.Source: §3 · p. 3
• The cloud maps an acquired context snapshot to waypoint guidance, while the edge maps current observations and available guidance to a trajectory.The trajectory contains ten points with components (x, y, sin ψ, cos ψ), executed by a separate locomotion controller.Source: §3 · p. 3
• At a local decision, the formulation selects among arrived cloud results using the newest source context; the chosen guidance may stay fixed as observations and trajectories change.The equations apply when usable guidance exists; startup, expiration, and communication loss require supervisory handling.Source: §3 · p. 3

4 ASYNCHRONOUS CLOUD–EDGE DUAL-POLICY NAVIGATION

• The cloud policy supplies route-level subgoals, while the onboard policy uses fresh RGB and LiDAR observations to generate obstacle-avoiding trajectories for a downstream locomotion controller.The two-stage training phases are not sequential inference steps.Source: §4 · p. 3

4.1 SHARED ROUTE INTERFACE

• Indoor planner paths and outdoor map routes share a route interface whose lookahead tokens connect the supplied route to visual observations and local subgoal guidance.The route window describes the upcoming segment in a local robot frame and carries source, availability, and reliability information.Source: §4.1 · p. 3

4.2 CLOUD AND EDGE NAVIGATION POLICIES

• The cloud model predicts subgoals from route tokens, instructions, and visual history; the edge fuses camera and LiDAR features to generate ten-point trajectories.The auxiliary VQA branch supervises traversability, road rules, region localization, and waypoint diagnosis; its performance is evaluated separately from navigation.Source: §4.2 · pp. 3, 5, Figure 1 · p. 4

4.3 ASYNCHRONOUS CONDITIONING AND EXECUTION

• At each edge decision, the latest arrived cloud guidance conditions a trajectory from fresh observations, while in-flight cloud results cannot affect earlier decisions.The schedule specifies dependencies, not independently verified deployment timing, guidance conversion, or fallback behavior.Source: §4.3 · p. 5, Figure 2 · p. 5

4.4 CONTINUOUS OPERATION AND SUPERVISION

• The supervisory design tracks route progress, trusted subgoals, failures, and visual anchors to support waiting, relocalization, replanning, or stopping.Recovery effectiveness requires episode-level evidence, and achieved update rates and communication latency require timing measurements.Source: §4.4 · p. 6

5 DATA AND TRAINING

• Stage 1 uses mixed-source, multi-task SFT and DAgger for cloud-policy training; Stage 2 uses primarily real-robot data for trajectory-VAE learning and flow-matching obstacle avoidance.These stages describe supervision and training sources, distinct from task-specific evaluation sets.Source: §5 · p. 6

5.1 NAVIGATION CORPUS

• The reported corpus contains over 650,000 clips, including 350,000+ real-world and 300,000+ simulation clips, and more than 1,000 hours of data.It includes 400,000+ outdoor and 250,000+ indoor clips; scene shares are reported separately.Source: §5.1 · p. 6, Figure 3 · p. 6

5.2 NAVIGATION IMITATION AND CORRECTIVE DATA

• Expert trajectories initialize navigation imitation, while DAgger-style aggregation adds expert labels at policy-visited states, including route deviations, incorrect turns, oscillations, and overshoot.The protocol reduces expert mixing across rounds and uses event-balanced recovery sampling.Source: §5.2 · pp. 6–7

5.3 GROUNDED VQA SUPERVISION

• Grounded VQA supervises road-rule understanding, traversable-class recognition, region localization, and waypoint diagnosis using answers associated with spatial evidence.The intended verification protocol checks claims using available depth, occupancy, maps, OCR, detections, and odometry, retaining uncertain cases as uncertain.Source: §5.3 · p. 7

6.1 EVALUATION SETTING AND PROTOCOL

• The main evaluation covers 258 episodes across 145 scenes under shared routes, rendered scenes, and closed-loop termination rules.Baseline information interfaces are not identical, and the evaluation describes a reported evaluator rather than verified physical-robot deployments.Source: §6.1 · p. 7

6.2 NAVIGATION RESULTS

• 55.04% SR and 55.03% SPL are the highest reported values, with gains of 37.21 and 37.20 percentage points over CityWalker, respectively.VbotNav-0 has lower non-traversable failure rates than CityWalker and SocialNav, but does not minimize every failure category.Source: §6.2 · pp. 7–8, Table 1 · p. 7

6.3 TRAINING ABLATIONS

• Removing DAgger data reduces SR by 17.01 percentage points and SPL by 17.05 points; removing VQA co-training reduces them by 9.23 and 9.35 points.The larger SR decrease occurs when DAgger data is removed.Source: §6.3 · p. 8, Table 2 · p. 8

6.4 GROUNDED VQA EVALUATION

• VQA training improves road-rule pass rate by 26.3 percentage points, traversable-class F1 by 46.7, localization F1 by 46.7, and waypoint-diagnosis pass rate by 41.5.The trained model's waypoint-diagnosis pass rate remains 45.0%; pass rates and F1 measure different properties.Source: §6.4 · p. 8, Table 3 · p. 9
• Qualitative VQA examples show crosswalk recovery, sidewalk separation, route-offset measurement, valid-route recognition, and road-rule response changes, alongside residual mistakes.Examples include an extra box in the crosswalk and a 5.81 m route offset against a 5.82 m reference.Source: §6.4 · p. 8

6.5 QUALITATIVE NAVIGATION AND FAILURE ANALYSIS

• Matched episodes show traversable initial goals in bridge, plaza, and stair examples, while later stalling or non-traversable predictions can still prevent episode success.Route snapshots also show a doorway passage, locally visible turns, and a Stop at the endpoint.Source: §6.5 · pp. 8–9

7 DISCUSSION AND LIMITATIONS

• Without a preplanned route, the current system cannot be deployed out of the box in unmapped or first-visit indoor environments.Mapping or route preparation is required; online exploration, spatial mapping, route generation, and guidance updates remain future work.Source: §7 · p. 9

8 CONCLUSION

• The conclusion frames VbotNav-0 as asynchronous cloud guidance coupled to edge-generated local trajectories, supported by navigation, ablation, and VQA evaluations.The evaluations provide evidence for task performance and the value of corrective and semantic supervision.Source: §8 · p. 9

A IMPLEMENTATION AND EVALUATION DETAILS

• The appendix records the available evaluation protocol and original qualitative evidence.Source: A IMPLEMENTATION AND EVALUATION DETAILS · p. 12

A.1 BASELINE INTERFACES

• CityWalker and SocialNav receive five causal RGB observations and a target from the visible route’s last point, which is transformed to each model’s coordinate convention.SocialNav retains its native navigation prompt and waypoint sampler.Source: §A.1 · p. 12
• NaVid and NaVILA receive route-derived language instructions and formatted image histories, with predicted navigation actions parsed into local goals.Source: §A.1 · p. 12
• ViNT receives an observation history and a goal-view preview rendered from the visible route, and its last predicted waypoint becomes the local goal.Native histories and ViNT’s goal preview remain inputs even when not displayed in matched-episode figures.Source: §A.1 · p. 12

A.2 GROUNDED VQA METRICS

• Traversable-class and localization F1 use micro-averaged matching, while road-rule and waypoint tasks use a fixed 0–4 judge rubric with scores of at least 3 passing.Localization requires class-consistent, one-to-one greedy box matching at IoU ≥0.50; semantic scoring penalizes unsupported claims, contradictions, and critical omissions.Source: §A.2 · p. 12

SPATIAL GROUNDING

• VQA training recovers a crosswalk and separates sidewalk regions, but the crosswalk example retains an extra box.The figure preserves the original boxes and IoU annotations.Source: Figure 4 · p. 13

WAYPOINT DIAGNOSIS

• Figure 5 presents waypoint diagnosis before and after VQA training, and its blue reference routes are shown only for evaluation.Source: Figure 5 · p. 14

A.4 MATCHED NAVIGATION CASES

• Selected bridge, plaza, and stair cases show VbotNav-0 choosing traversable first goals; baseline first-goal and later-rollout failures vary across cases.These selected cases are not an estimate of failure frequency.Source: Figure 7 · p. 16, Figure 8 · p. 16, §A.4 · p. 16

A.5 INDOOR–OUTDOOR ROLLOUT SNAPSHOTS

• Snapshots trace doorway passage, locally actionable left and right turns, and stopping at the route endpoint; displayed rationale does not verify safe execution.The panels pair predicted local trajectories and rationale with route progress and pose.Source: §A.5 · p. 17, Figure 10 · p. 18