PaperlayerView exact revision on arXiv

Source-linked AI summary

WAVE-Go: World-Model Navigation with Adaptive Execution for Wheel-Legged Robots

Mingyi Li, Ji Li, Zhihao Ouyang, Yage He, Börje F. Karlsson

arXiv:2609.18193v1cs.RO

TL;DR

Wheel-legged robots need navigation execution that can adapt when predicted action sequences become invalid because of obstacles or locomotion-mode changes. WAVE-Go selects and interrupts action prefixes using estimated cumulative risk and validity checks, achieving a 7.7% dynamic-OOD success gain over the strongest baseline while reducing collisions.

  • Problem

    Wheel-legged navigation lacks execution strategies that balance long-horizon planning efficiency with frequent revision under changing obstacles and locomotion-mode constraints.

  • Method

    WAVE-Go selects action prefixes under a cumulative failure budget and interrupts pending commands when observations or mode-validity checks invalidate execution.

  • Results

    WAVE-Go achieves a 7.7% absolute dynamic-OOD success gain over the strongest baseline and a 34.1% reduction in collisions per 100 m.

  • Takeaways & Limitations

    Adaptive execution improves navigation across evaluated scene conditions while balancing performance against replanning overhead.

Abstract

from arXiv · show

World models can anticipate the consequences of navigation actions, but predicted action sequences may become invalid during execution, especially when wheel-legged robots encounter dynamic obstacles or change locomotion modes. We propose WAVE-Go, an image-goal navigation framework that separates world-action prediction from interruptible command execution. Its executor adaptively selects an action prefix and cancels pending commands when updated observations invalidate execution. A conditional-risk formulation specifies prefix selection under an estimated cumulative failure budget, while posture and locomotion-mode transitions require clearance, stability, and task-evidence checks. In the reported navigation evaluation, WAVE-Go achieves 74.1% in-distribution success and 63.3% dynamic out-of-distribution success, exceeding the strongest baseline by 4.7 and 7.7 percentage points, respectively, while reducing collisions from 4.4 to 2.9 per 100 m. Compared with interruptible fixed four-command execution, WAVE-Go raises success by 4.0 percentage points while reducing replanning frequency by 51.2% and collision rate by 6.5%. Execution ablations also show that runtime interruption improves success, collision rate, and reaction latency at the cost of additional replanning. These results support adaptive, interruptible execution as a means of balancing navigation performance and planning overhead. Code is available at https://github.com/vigorlee/wave-go.

I. INTRODUCTION · II. RELATED WORK

WAVE-Go addresses the mismatch between predicted navigation actions and changing execution conditions by combining joint world–action prediction with adaptive, interruptible, mode-aware control. It builds on world models, closed-loop policies, action chunking, and transition checks while targeting a cumulative-risk trade-off between commitment, replanning, and responsiveness.

  • I. INTRODUCTION: Wheel-legged execution must account for changing geometry, target occlusion, posture transitions, and locomotion-mode dependencies that can invalidate predicted commands.Long chunks reduce planning frequency but increase commitment, whereas short chunks enable revision at greater planning cost.
  • I. INTRODUCTION: WAVE-Go couples goal-conditioned joint world–action prediction with adaptive, interruptible execution for wheel-legged image-goal navigation.The predictor uses goals, observations, robot motion, and structured task state to generate candidate actions and future-state features.
  • I. INTRODUCTION: The executor selects the longest valid prefix from 4, 8, and 16 commands under an estimated cumulative failure budget before replanning.This formulation combines conditional failure estimates with validity checks to authorize execution length.
  • I. INTRODUCTION: Pending commands are revalidated and interrupted when needed, with posture and locomotion-mode transitions treated as boundaries requiring configuration confirmation.Transition checks incorporate clearance, stability, and task readiness before subsequent motion.
  • I. INTRODUCTION: WAVE-Go targets a performance–cost trade-off by improving navigation with fewer replanning calls than short fixed-prefix execution, while interruption adds replanning cost for greater responsiveness.The framework balances commitment to predictions against changing conditions rather than relying on predictive uncertainty alone.
  • II. RELATED WORK: Prior world-model navigation methods predict future observations or visual features for trajectory optimization and goal-directed action selection, including NWM, DINO-WM, and related rollout approaches [8]–[10].These methods address representation, prediction cost, and temporal consistency.
  • II. RELATED WORK: Joint world–action models include FutureNav [4], NavWAM [5], WAM-Nav [6], NavWM [14], WorldVLN [15], while other work studies closed-loop policies [17], embodiment transfer [18], and action chunking.Adaptive chunking uses predictive uncertainty [7], whereas WAVE-Go additionally emphasizes interruptibility and mode-aware execution.

III. TASK FORMULATION

WAVE-Go formulates image-goal navigation as a partially observed control problem with multimodal observations, goal and state inputs, and locomotion mode. Its joint world-action rollout supports adaptive execution that selects a command prefix and cancels pending commands when updated observations invalidate the plan.

  • III. TASK FORMULATION: Navigation is modeled as a partially observed control problem using RGB, depth, local range, goal-image, optional language, proprioceptive, and locomotion-mode inputs.Updated observations are required because history cannot fully resolve occlusions or future scene changes.
  • III. TASK FORMULATION: The executor selects at most K_t ≤ H commands and cancels pending commands when updated observations invalidate execution.This separates joint rollout prediction from interruptible command execution.
  • III. TASK FORMULATION: The semantic task state records goal progress and dependencies, while a motion-context token encodes recent visual motion, proprioception, actions, and optional tracking features.The history includes preceding observations and intervening actions but excludes the action selected at the current time.
  • III. TASK FORMULATION: Each of H = 16 predicted steps pairs a candidate action and locomotion mode with a successor latent state, while auxiliary features provide uncertainty and command-validity cues.The interface predicts candidate futures without conditioning on an externally prescribed action sequence.

IV. METHOD · A. Predictor Adaptation

WAVE-Go adapts joint world–action prediction to wheel-legged embodiment and connects it to adaptive, interruptible execution. Its post-training data and supervision emphasize temporally aligned transitions, observable failures, and events that challenge navigation and locomotion.

  • IV. METHOD: WAVE-Go uses calibrated conditional-risk prefix selection, observation-triggered command interruption, and posture or locomotion-mode verification to connect prediction with execution.Posture or mode events require configuration confirmation and a new prediction.
  • IV. METHOD: The architecture jointly predicts candidate actions, successor latent states, locomotion modes, and auxiliary cues from multimodal context.These predictions support conditional-risk prefix selection and interruptible execution.
  • A. Predictor Adaptation: Predictor adaptation uses temporally aligned observation–action transitions linking histories to executed actions, locomotion modes, and resulting observations.The post-training formulation is built around these aligned transitions.
  • A. Predictor Adaptation: Goal images and terminal-event annotations distinguish visually similar locations from task completion, while failed observable transitions supervise consequence prediction and failure estimation.The labels capture whether navigation reaches the task-defined terminal event rather than only a visually similar location.
  • A. Predictor Adaptation: The baselines’ visual–language–action pairs are adapted to wheel-legged transitions involving occlusions, dynamic obstacles, target reacquisition, posture changes, and locomotion-mode transitions.This targets the observation and action distribution of the wheel-legged platform; the frozen-model trace does not measure the curriculum’s benefit.
  • A. Predictor Adaptation: Transition sampling is stratified around target discovery, occlusion, obstacle appearance, stopping, and mode changes to reduce forward-motion dominance.Matched successful and failed transitions provide additional supervision where available.

B. Multimodal Context · C. World–Action Learning

WAVE-Go fuses multimodal goal, observation, language, task-state, capability, and motion information into a structured representation for world–action prediction. Its joint objective learns future states, inverse dynamics, actions, and locomotion modes while enforcing action–state consistency.

  • B. Multimodal Context: The semantic reasoner combines the goal image, selected observations, optional language, previous task state, and available robot capabilities.Its output follows a fixed schema for downstream prediction.
  • B. Multimodal Context: Qwen-3-VL produces structured state variables for goal progress, accumulated evidence, subgoals, confidence, and ordered dependencies without generating high-rate velocity commands [22].These variables provide task-relevant semantic conditioning for the world–action head.
  • B. Multimodal Context: The fused representation integrates visual observations, the navigation goal, recent robot motion, optional language, and motion tokens.A learned null token replaces unavailable language input, while auxiliary objectives supervise proprioception, short-range egomotion, target displacement, and action–state-transition consistency.
  • B. Multimodal Context: The world–action head is conditioned on the fixed-schema task state, which encodes task-relevant information from semantic reasoning.This connects multimodal semantic context to predictive action modeling.
  • B. Multimodal Context: Stop-gradient operations limit goal-identity encoding in the motion token, while motion and locomotion-mode probes test alignment with observation-history information.A shuffled-motion-token control assesses whether the representation captures motion information rather than spurious associations.
  • C. World–Action Learning: The predictive objective jointly supervises future latent-state prediction, inverse dynamics, action generation, and locomotion-mode prediction.Its consistency term encourages predicted actions to agree with their associated future states, building on joint world–action formulations [4], [5].

D. Hybrid Commands · E. Conditional-Risk Prefix Selection

The executor integrates continuous drive and posture commands with discrete locomotion-mode events, using validity masks and configuration-dependent prefix boundaries. It selects the longest admissible prefix under a calibrated conditional-risk budget, while runtime checks can cancel commands and trigger replanning.

  • D. Hybrid Commands: Hybrid commands combine continuous drive and posture controls with discrete locomotion-mode events under controller-specific validity masks.Posture and mode transitions end the current execution prefix because subsequent driving must reflect the resulting robot configuration.
  • E. Conditional-Risk Prefix Selection: Longer prefixes reduce planning frequency but prolong execution without updated world–action predictions [7].The method therefore estimates cumulative failure risk over candidate prefixes to balance planning frequency against continued-execution risk.
  • E. Conditional-Risk Prefix Selection: The risk model combines world-model disagreement, action entropy, obstacle proximity, dynamic-scene evidence, mode status, and sensor age.Conditional survival probabilities produce a cumulative failure estimate for each candidate prefix.
  • E. Conditional-Risk Prefix Selection: Among K = {4, 8, 16}, the executor selects the longest prefix satisfying the estimated risk budget δ and validity test G_t(k).The formulation accounts for survival through preceding commands without assuming independent failures.
  • E. Conditional-Risk Prefix Selection: Validity checks enforce command validity, mode preconditions, and observation freshness; when none is admissible, K_t = 0 triggers stopping and renewed sensing.
  • E. Conditional-Risk Prefix Selection: Runtime checks cancel remaining commands and trigger replanning when clearance, sensing freshness, task state, or prediction residuals become unacceptable.Posture and mode events also terminate prefixes, requiring configuration confirmation and a new prediction before subsequent driving.
  • E. Conditional-Risk Prefix Selection: Calibration evaluates cumulative risk at every candidate length on a separate scene split because stepwise calibration does not guarantee prefix-level calibration.The estimated risk depends on learned conditional probabilities, so δ is an estimated failure budget rather than a certified physical collision bound.

F. Evidence-Gated Hybrid Transitions

WAVE-Go gates posture and locomotion-mode transitions using state, safety, task-evidence, and controller checks before dispatch. The executor separately handles rejected commands and records interventions to distinguish them from execution failures.

  • Transition prerequisites: Transition gates require admissible mode changes plus freshness, stability, clearance, goal evidence, and controller validity before dispatch.Crouching additionally requires the robot to be stationary and stable with sufficient obstacle clearance, while task-specific transitions depend on goal evidence and prior dependencies.
  • Transition validation: Rejected proposals are checked independently before execution, so successful interventions are not counted as execution failures.The gate contributes to the prefix-validity test G_t(k), and rejected transitions are recorded separately from transitions that fail after dispatch.
  • Invalid-command handling: Invalid commands trigger type-specific responses: rejected drives renew sensing, rejected posture or mode tokens select no substitute, and malformed semantic states suspend progression.When semantic state is malformed, the executor retains the last validated state.
  • Execution records: Execution records link outcomes with predictions, risk estimates, and runtime decisions to distinguish prediction errors, interventions, and command-tracking failures.Rejected commands are logged separately because no observed successor state can be attributed to their execution.

G. Multi-Rate Execution

WAVE-Go separates feedback frequency from prediction frequency: tracking and validity checks run continuously, while world-action prediction and semantic reasoning update only when execution or task conditions require it. Buffered predictions are reused only after freshness, mode, and task-state checks, while deployment metrics must distinguish controller frequency, inference latency, and observation-to-command latency.

  • Multi-Rate Execution: Multi-rate scheduling runs command tracking and validity checks at controller frequency, triggers world-action inference when prefixes expire or are cancelled, and updates semantic reasoning when task evidence changes.Buffered predictions are checked against observation freshness, current mode, and task state before reuse.
  • Multi-Rate Execution: Separating feedback from prediction frequency allows longer execution prefixes to reduce model invocations while preserving checks before every command.New observations can invalidate pending actions without waiting for prefix expiration.
  • Multi-Rate Execution: Mode transitions trigger renewed prediction because they change the conditions governing subsequent motion, separating continuation decisions from new action-chunk generation.This scheduling strategy addresses the tradeoff between timely feedback and repeated inference.
  • Multi-Rate Execution: Deployment evaluation should separately measure model inference latency, observation-to-command latency, memory use, numerical agreement, and controller frequency because reported action-selection rate includes task-level processing.TensorRT acceleration and reasoning-model quantization are deployment options requiring separate measurement.

V. EVALUATION · A. Experimental Setup

The evaluation tests whether WAVE-Go improves closed-loop navigation across in-distribution and out-of-distribution conditions, and examines how execution policies trade task performance against replanning and response behavior. The setup uses controlled comparisons, explicit navigation metrics, and separate measures for collisions, mode transitions, replanning, reaction latency, and command variation.

  • V. EVALUATION: The evaluation asks whether WAVE-Go improves closed-loop navigation across scene conditions and how prefix selection and interruption affect performance–replanning trade-offs.Quantitative results, execution-policy ablations, qualitative sequences, and additional action-selection measurements address these questions.
  • A. Experimental Setup: Comparisons cover in-distribution, scene-OOD, and dynamic-OOD navigation using action-only, NWM, FutureNav-style [4], and NavWAM-style [5] baselines.The conditions represent familiar scenes, unseen scenes, and changes in dynamic obstacles.
  • A. Experimental Setup: Scene-disjoint splits and paired episodes hold initial poses, goals, obstacle schedules, sensors, command limits, and timeouts constant across comparisons.Execution-policy variants share a frozen predictor; full WAVE-Go and no-interruption use the same conditional-risk model, differing only in command cancellation.
  • A. Experimental Setup: Navigation success requires reaching the goal region within timeout while satisfying any terminal-mode requirement; internal completion prediction alone is insufficient.Success rate is computed from binary episode outcomes, and SPL uses success weighted by path length [24].
  • A. Experimental Setup: SPL uses traversability-constrained shortest feasible paths, assigns failed episodes zero, penalizes successful detours, and is omitted when no valid reference path exists.Completion time can instead characterize delays caused by dynamic obstacles; tables report SR and SPL as percentages.
  • A. Experimental Setup: Collision rate is distance-normalized and includes failed episodes, while mode-transition reporting distinguishes rejected proposals from dispatched transitions that fail preconditions.Continuous contact counts once until separation, excluding intended ground and docking contacts; collision rate should be interpreted alongside success.
  • A. Experimental Setup: Replanning frequency counts model invocations per 100 m, whereas reaction latency measures the 95th-percentile delay from invalidating observations to stop or replacement commands.These metrics distinguish planning frequency from inference time and delayed command responses; command variation is measured separately from normalized drive commands.

B. Quantitative Results

WAVE-Go achieves the strongest reported success rates across all three scene conditions and the highest in-distribution SPL, while reducing collisions relative to the strongest baseline. Its larger dynamic-OOD gain is consistent with adaptive interruption limiting outdated action execution, though collision rates are reported only as distance-normalized totals.

  • Success across scene conditions: 4.7%, 5.1%, and 7.7% absolute SR gains over the strongest evaluated baseline in ID, scene-OOD, and dynamic-OOD, respectively, with the highest reported SR in all conditions and 5.0% higher ID SPL.Table I also reports WAVE-Go as having the highest ID SPL.
  • Success across scene conditions: The larger dynamic-OOD improvement is consistent with adaptive prefix selection and interruption reducing continued execution of outdated action sequences when obstacles move or execution conditions change.The supplied passages connect changing obstacles or execution conditions to the need to stop outdated sequences, but do not provide separate mechanism-level measurements.
  • Collisions and task completion: 34.1% lower collision rate than the strongest baseline, alongside higher task success rather than collision reduction from simply stopping.The comparison uses one distance-normalized collision rate per method, not separate rates for each scene condition.

C. Execution-Policy Ablations

Execution-policy ablations show that WAVE-Go’s adaptive selection improves navigation while reducing replanning relative to fixed prefixes, and that combining predictive and execution cues outperforms single-cue selection. Runtime interruption further improves success, collision rate, and responsiveness, while fixed longer prefixes reduce replanning but harm navigation.

  • Fixed execution lengths: 70.1% to 66.0% success as K increases from 4 to 16, while replanning falls from 48.6 to 17.8 invocations per 100 m and collisions rise from 3.1 to 6.7 per 100 m.Longer fixed prefixes also increase reaction latency, showing that fewer planning calls and smoother commands do not necessarily improve navigation.
  • Adaptive selection: 4.0% higher success and 51.2% fewer replanning calls than fixed K = 4, with fewer collisions per 100 m (2.9 versus 3.1).The trade-off is a 0.03 s increase in reaction p95, from 0.18 to 0.21 s.
  • Adaptive selection: 2.2% higher success and 14.7% fewer collisions than world-uncertainty-only selection, while also reducing replanning frequency and reaction p95.Combining predictive cues with execution conditions better captures whether a command remains executable than uncertainty alone.
  • Runtime interruption: 3.6% higher success, 45.3% fewer collisions per 100 m, and 61.1% lower reaction p95 with runtime interruption, at the cost of 17.9% more replanning calls per 100 m.Cancelling pending commands when updated observations invalidate execution improves task completion and responsiveness while reducing collisions.

D. Qualitative Results · E. Additional Measurements · VI. CONCLUSION

WAVE-Go combines budgeted adaptive prefix selection with interruptible, evidence-gated execution for wheel-legged navigation. Across qualitative demonstrations and measurements, it supports posture transitions, prompt command cancellation, improved navigation outcomes, and reduced replanning relative to fixed-horizon execution.

  • D. Qualitative Results: Qualitative simulations cover Go2-W household search, LUNA corridor navigation, continuous LUNA physics rollout through red pillars, and Go2-W stair and obstacle traversal.Narrow passages require checking whether upcoming commands remain feasible as obstacles approach.
  • D. Qualitative Results: The real-world experiment shows the robot approaching and aligning with a station before lowering its posture, demonstrating navigation-to-task transition under a required configuration.The transition illustrates task-stage execution that depends on a specific robot posture.
  • E. Additional Measurements: The Go2-W long-horizon measurement lasted 297.891 s over 121 action-selection rounds, with a mean selection rate of 0.4062 Hz using Cosmos3-Edge [25] on an NVIDIA L20s GPU.Table III reports the action-selection measurements, while Figure 4 provides separate LUNA trials.
  • E. Additional Measurements: Separate LUNA trials distinguish command cancellation from subsequent body settling and show that pending commands can be cancelled without waiting for the selected prefix to finish.Cancellation remains prompt even with longer execution prefixes.
  • VI. CONCLUSION: WAVE-Go selects action prefixes under a cumulative failure budget and validity constraints, cancels pending commands after invalidating observations, and gates posture or locomotion transitions on configuration evidence and renewed prediction.These mechanisms connect world-action prediction to adaptive, interruptible execution for wheel-legged robots.
  • VI. CONCLUSION: 7.7% higher dynamic-OOD success and 34.1% fewer collisions per 100 m than the strongest baseline, while short fixed-horizon execution trails by 4.0% success and 51.2% more replanning calls.Runtime interruption further improves success, collision rate, and reaction latency, at the cost of additional replanning.
Loading 2609.18193v1…