PaperlayerView exact revision on arXiv

Source-linked AI summary

Future-Aware Flow Planning for Safe UAV Target Following

Boning Feng, Haoran Zhang, Xiaowen Bi, Yanzhen Zhang, Xiaodan Shi

arXiv:2609.10166v1cs.RO

TL;DR

UAV target following in clutter requires anticipating target motion without committing the UAV to unsafe or blocked target paths. The paper proposes future-aware flow planning with horizon-aligned residual future guidance and embedded executable-prefix repair. On fixed ID/OOD benchmarks, it improves the targeted safety–tracking trade-off, while Future-MPC remains better on some smoothness and thresholded success measures.

  • Problem

    Current-state followers can lag behind target turns or choose unsafe corridors, while direct target-path following may copy blocked or narrow routes.

  • Method

    The method separates target anticipation from UAV route generation by using predicted futures as residual flow guidance and embedding RSEPSS prefix repair during sampling.

  • Results

    The planner achieves zero measured ID collision rate with the highest ID STT@8, and the lowest OOD collision rate and best average FDE among displayed methods.

  • Takeaways & Limitations

    Horizon-aligned future adaptation and embedded prefix repair are complementary ingredients for safety-constrained UAV target following under the tested simulation conditions.

  • Takeaways & Limitations

    The evaluation does not replace onboard perception or physical flight validation and excludes dynamic obstacles, end-to-end visual detection, re-identification, and hardware-specific aerodynamics.

Abstract

from arXiv · show

UAV target following in cluttered environments is inherently predictive: current-state followers can lag behind turns, choose blocked corridors, or trade tracking for unsafe near-horizon motion. We propose a future-aware flow planning framework for state-informed UAV target following. Predicted target futures guide clean UAV trajectory generation as horizon-aligned residual signals, while risk-scored executable-prefix repair is embedded inside the sampling loop. On fixed ID/OOD receding-horizon benchmarks, the planner improves the intended safety--tracking trade-off rather than dominating every metric: it matches zero measured ID collision rate with the highest ID safe-tracking time, and gives the lowest OOD macro collision rate and final tracking error among the displayed methods, while Future-MPC remains smoother and stronger on some thresholded OOD success metrics under its hand-designed objective. Ablations show that future adaptation improves candidate generation before safety repair, and simulator-facing stress tests probe interface, sensing, and controller-execution effects. These results support horizon-aligned future adaptation and embedded prefix repair as complementary ingredients for safe UAV target following under the tested simulation conditions.

1 Introduction

UAV target following requires predictive, receding-horizon planning that converts uncertain target futures into UAV-feasible trajectories while preserving tracking and executable safety. Future-Aware Flow Planning addresses this by combining horizon-aligned future guidance with embedded prefix repair, improving the targeted safety–tracking trade-off on fixed benchmarks.

  • The planner receives discrete target-state observations, the UAV state, and ego-frame local obstacles, executes a short prefix, and then replans.
  • Current-state following can lag behind turns, commit to soon-obsolete corridors, or react late when clutter restricts feasible routes.
  • Directly using predicted target paths as UAV references is unsafe because targets may traverse narrow or blocked regions.
  • Future-Aware Flow Planning separates target anticipation from UAV route generation by injecting predicted futures as horizon-aligned residual guidance.
  • RSEPSS operates inside each sampling update so risky near-horizon executable segments influence trajectory generation rather than being repaired only afterward.
  • On fixed ID/OOD benchmarks, the method reaches zero measured collision rate and highest ID STT@8, plus lowest OOD collision rate and best average FDE among displayed methods.

2 Related Work

Prior work spans optimization-based tracking, generative trajectory planning, and safety correction, but the proposed framework combines future-conditioned flow generation with correction focused on the executable prefix. Its design keeps future motion as guidance while preserving a UAV-feasible generated trajectory.

  • Classical tracking and MPC-style planners use predicted target motion, while quadrotor replanning emphasizes perception-aware local planning, feasibility, and dynamics.
  • Generative planners represent multimodal trajectories, with flow matching offering continuous transport dynamics for forecasting, generation, and robot planning.
  • The framework uses clean-trajectory x-pred flow generation because UAV target following outputs the executable trajectory itself.
  • Gated residual future tokens address the tension between useful structured future signals and unreliable references that can mislead tracking.
  • Unlike corrections applied to full trajectories or full sampling dynamics, RSEPSS focuses safety correction on risky prefix points because only a short prefix executes before replanning.

3 Methodology

The framework combines future-aware clean-trajectory flow generation with embedded repair of risky executable prefixes. Target futures guide UAV-feasible trajectories, while RSEPSS modifies prefix candidates during sampling so safety correction influences subsequent flow updates.

  • Future-aware trajectory generation: The planner predicts UAV trajectories directly, using target-motion hypotheses to resolve horizon-level tracking ambiguity without copying the target path.The generated object remains a UAV-feasible trajectory, allowing obstacle-aware detours.
  • Conditional flow planner: A future encoder supplies temporal target tokens alongside obstacle geometry and motion context to the Flow Transformer planner.Trajectory, flow-time, and horizon-position embeddings represent intermediate samples before conditional generation.
  • Gated future adaptation: The gated future-adapter injects horizon-aligned target information as a late residual refinement whose strength is controlled by a learned reliability gate.This preserves the base planner’s obstacle-aware route structure while changing tracking intent.
  • Executable-prefix safety repair: RSEPSS scores executable-prefix clearance with larger weights on earlier points, then repairs high-risk points and temporal neighbors by projecting them away from nearby obstacles.A smoothing step reduces abrupt local changes while keeping the prefix start consistent with the current UAV state.
  • Embedded sampling: The corrected clean candidate is fed back into the flow sampler at every sampling step, rather than applied only as a detached post-sampling filter.Only the executable prefix is modified, leaving the longer horizon available for target-aware replanning.

4 Experiments

Experiments evaluate future-aware generation, embedded prefix repair, robustness to distribution shift, and simulator-facing execution. The planner improves the safety–tracking trade-off, while Future-MPC remains stronger on selected smoothness and thresholded success metrics.

  • Evaluation design: The fixed ID/OOD benchmark tests future-aware generation and embedded prefix repair against safety-aware learned flow baselines and Future-MPC.All methods share the horizon, executed prefix, obstacle budget, target predictor, and metrics.
  • ID benchmark: 0.000 collision rate and 0.960 STT@8 make the proposed planner the safest displayed method on the nominal ID benchmark.Future-MPC reports CR 0.043 and STT@8 0.583 in dense clutter, despite lower ADE/FDE and jerk.
  • ID benchmark: The planner accepts endpoint-error cost to preserve an executable safe prefix near obstacles.This trade-off is visible in the closed-loop benchmark comparison.
  • OOD benchmark: 0.002 collision rate and 3.719 average FDE are the proposed method’s best displayed OOD macro results.Future-MPC favors lower jerk and thresholded STT@8/STR@8, but does not dominate tracking accuracy or safety.
  • Simulator-facing validation: PX4 SITL validation uses 120 offboard executions across 60 ID forest-medium and 60 OOD combined-shift wall scenes.References are streamed to PX4 at 20 Hz using 36 points per run.
  • Ablations: The gated future adapter reduces ID ADE/FDE from 4.628/5.032 to 3.462/4.453 and OOD ADE/FDE from 5.243/4.822 to 3.002/2.698 before RSEPSS.The ablation isolates future-conditioned candidate generation under matched predictor, interface, training, and rollout settings.
  • Ablations: With an 8 m biased future, RSEPSS limits STT@8 degradation to −0.004 ID and −0.012 OOD while keeping collision rate essentially unchanged.This supports gated residual guidance rather than treating predicted futures as hard trajectory anchors.

5 Conclusion and Limitations

The paper separates target prediction from UAV execution by using bounded future residuals and executable-prefix repair. Its conclusions are limited to tested simulation conditions and exclude onboard perception, physical flight, dynamic obstacles, and hardware-specific aerodynamics.

  • Conclusion: Future-Aware Flow Planning uses horizon-aligned residual target guidance for UAV trajectory generation, while RSEPSS repairs risky executable-prefix points during sampling.The generated object remains a UAV-feasible trajectory rather than a copied target path.
  • Conclusion: Future-MPC is a complementary hand-designed reference, not a method that one approach universally dominates.The reported results support a safety–tracking trade-off rather than universal metric superiority.
  • Limitations: The evidence remains simulation-bound: PX4 SITL does not replace onboard perception or physical flight validation, and the framework excludes dynamic obstacles and hardware-specific aerodynamics.It also assumes discrete target-state observations and local obstacle inputs.
  • Limitations: RSEPSS improves executable safety but increases replanning cost, motivating lighter repair, distillation, or adaptive triggering for onboard deployment.This is identified as a deployment constraint of the safety mechanism.

A.1 Benchmark, Baselines, and Simulator-Facing Protocol

The evaluation uses fixed ID/OOD receding-horizon benchmarks with matched baselines, metrics, and execution settings, complemented by simulator-facing and controller-in-the-loop tests. Metrics cover tracking, collisions, smoothness, safe tracking, success, and replanning runtime under a fixed rollout protocol.

  • Benchmarks: The ID benchmark uses forest scenes across low, medium, and high obstacle densities, while OOD varies obstacle shape, distribution, target speed, or their combination.
  • Baselines: Future-MPC shares the target predictor and rollout protocol but uses a hand-designed receding-horizon objective with cylinder-geometry clearance.
  • Metrics: The main tables report tracking accuracy, collision rate, jerk-RMS smoothness, safe tracking time, thresholded success, and seconds per replan.
  • Simulator-facing validation: Additional tests perturb target-state, point-cloud, and latency interfaces, use Isaac-style LiDAR sensing, and validate execution through PX4 flight-controller-in-the-loop runs.
  • Metric definitions: Collision rate is the mean dense hard-collision rate over executed rollout states, capturing how often the closed-loop prefix enters collision.
  • Metric definitions: Safe tracking time measures the fraction of rollout time satisfying the tracking criterion, while safe tracking success requires collision-free completion within the tolerance.
  • Rollout protocol: The fixed rollout protocol uses horizon H = 64, executed prefix length Kexec = 4, 12 flow updates per replan, and a maximum of 25 replans.

B Implementation and Training Details

The planners use a Transformer-style trajectory backbone and Mamba-based target forecaster, trained on scene-aligned successful expert records with controlled data splits. Results and uncertainty are reported through fixed-case tables and bootstrap summaries.

  • Architecture: All learned planners use a Transformer-style trajectory backbone with dmodel = 256, 6 layers, 8 attention heads, 128 obstacle tokens, and horizon length 64.
  • Uncertainty reporting: Table 5 reports bootstrap uncertainty for claim-critical benchmark metrics, while Table 6 reports bootstrap uncertainty for PX4 flight-controller-in-the-loop validation.
  • Architecture: The target-motion predictor is Mamba-based, using history length 16, horizon length 64, and 4 selected future hypotheses.
  • Training data: Training uses successful expert samples from scene-aligned JSONL files split by scene id into 80% training, 10% validation, and 10% test partitions.
  • Adapter training: The reported adapter run uses 200 scene-aligned training files, batch size 8, AdamW, learning rate 2 × 10^-4, weight decay 10^-4, gradient clipping at 1.0, and 25 epochs.

C Additional Predictor and Benchmark Results

Predictor evaluation separates open-loop future forecasting from downstream planner behavior. The evidence treats predicted futures as useful but uncertain guidance, not as hard UAV route anchors.

  • Predictor quality: The open-loop predictor probe evaluates predicted target futures against ground truth on held-out records across trajectory-complexity bins.
  • Predictor quality: Table 7 reports open-loop target predictor quality, with complexity defined by the trajectory-complexity score and finite rate measuring numerically valid predictions.
  • Robustness: The predictor outputs are passed as uncertain feature guidance, and a biased-future stress test measures planner behavior when predicted futures are wrong.
  • Ablations: Fixed-case ablations show that predicted target futures improve target-following accuracy under a matched rollout protocol, while future paths are not treated as hard trajectory anchors.
  • Benchmark results: Tables 8 and 9 provide full group-wise ID and OOD results, with smoothness measured by jerk RMS.

D OOD Benchmark Design and Qualitative Results

The OOD benchmark separates deployment mismatches by interpretable shift types and uses matched visualizations to compare failure modes. Qualitative results highlight different safety, tracking, and computational trade-offs among the displayed methods.

  • OOD design: The OOD benchmark uses four controlled shift types to distinguish deployment-time mismatches rather than reporting one aggregated robustness number.
  • OOD design: Shape and distribution shifts alter obstacle conditions, speed shift alters target-future conditions, and the combined shift couples both sources of mismatch.
  • OOD results: Table 9 reports full OOD comparisons by shift type, with smoothness measured by jerk RMS.
  • Qualitative comparison: Figure 6 holds scene, obstacle rendering, target trajectory, axis bounds, and method ordering constant across displayed OOD rollouts.
  • Qualitative comparison: SafeFlow improves safety through stronger sampling-time correction but can become more conservative, whereas SafeFlowMatcher is computationally light yet more prone to weakened tracking in difficult shifts.
  • Qualitative comparison: Future-MPC is a fixed optimization-based reference, while the proposed method emphasizes learned generation with local prefix safety in several shifted scenes.

E Simulator-Facing Stress Tests

Simulator-facing tests examine whether the final planner remains stable under degraded interfaces and sensing. Across interface perturbations and sensor-derived observations, performance remains robust in the tested settings, with combined OOD shifts hardest.

  • L1: Interface perturbations: Across 682 records, the forest/medium setting remains collision-free under all tested interface stresses.The stresses include target-state perturbations, degraded obstacle observations, latency up to 300 ms, and combined mild stress.
  • L1: Interface perturbations: STT@8 stays above 0.93 under 300 ms latency and combined mild stress in the forest/medium setting.
  • L1: Interface perturbations: Under harder OOD combined shifts, performance degrades but safe tracking remains non-zero across all tested interface perturbation profiles.
  • L2: Isaac-style sensing: Under sensor-derived observations, the method preserves the forest trend, while the OOD combined-shift setting remains the most difficult case.

F Robustness Analysis under OOD Shifts

The robustness analysis is conditional: it decomposes OOD task risk under boundedness and Lipschitz assumptions, linking future-token reliability, clean-trajectory prediction, and embedded prefix repair. The resulting interpretation is a safety–tracking trade-off rather than universal safety dominance.

  • Scope and assumptions: The OOD robustness analysis is conditional and does not claim a universal safety certificate.It assumes standard boundedness and Lipschitz conditions for the evaluated setting.
  • Notation and assumptions: The observed planning condition c includes UAV state, target history, and local obstacle observations.
  • Notation and assumptions: The executed candidate is obtained after applying the embedded RSEPSS repair operator with margin m.
  • Notation and assumptions: The task loss is assumed locally Lipschitz near evaluated rollouts, which is weaker than requiring global robustness over arbitrary scenes.
  • Future-token influence: Reliability-gated residual adaptation mediates future-prediction corruption through residual scale, gating, and feature-level attention rather than hard trajectory anchoring.
  • Clean-trajectory prediction: Clean-trajectory x-pred directly controls task-risk deviation, whereas auxiliary parameterizations mediate feasible-trajectory deviation through reconstruction.
  • Prefix repair: Embedded repair lets safer prefixes influence subsequent sampling states while keeping the correction local and bounded.
  • OOD interpretation: Under the hardest combined shift, the method may trade safe-tracking success against the MPC reference while retaining low collision and final-error behavior.

G.1 Embedded RSEPSS Ablation

The ablations separate the contribution of embedded RSEPSS from clean-trajectory parameterization. Embedded repair substantially improves safe tracking by changing sampling states, while x-pred improves tracking and safe-tracking success relative to v-pred, at a smoothness cost.

  • Embedded RSEPSS: RSEPSS raises safe tracking from 0.282 to 0.913 in ID and from 0.646 to 0.854 in OOD.The embedded-RSEPSS rows remove nearly all dense collisions relative to the no-RSEPSS rows.
  • Embedded RSEPSS: Embedded RSEPSS moderately increases ADE/FDE and replanning time because refinement prioritizes executable safe prefixes over unconstrained geometric tracking.
  • Embedded RSEPSS: Applying RSEPSS inside sampling changes subsequent denoising states rather than merely clipping the final trajectory.The corrected prefix therefore influences later flow updates.
  • x-pred versus v-pred: The x-pred parameterization substantially reduces tracking error and improves safe-tracking success, especially under OOD shifts, relative to v-pred.Both variants use the same predictor conditioning and basic sampling interface without RSEPSS or the gated future adapter.
  • x-pred versus v-pred: The v-pred baseline produces lower jerk, but its smoothness is coupled to poor tracking and low success.
  • x-pred versus v-pred: The x-pred ablation supports clean-trajectory design but does not claim that x-pred alone solves safety.
Loading 2609.10166v1…