PaperlayerView exact revision on arXiv

Source-linked AI summary

PATH: Preference-Aware Traversability Estimation for Robotic Teleoperation in Off-Road Navigation

Xianghao Meng, Wenshuo Wang, Cheng Shao, Lewei Feng, Qingxiao Liu, Canyu Wang, Junqiang Xi

arXiv:2610.30605v1cs.ROcs.HC

TL;DR

Off-road teleoperation methods often overlook operator preferences for risk and speed, although those preferences shape route and speed choices. PATH-Net predicts candidate cost and speed maps from terrain features and vehicle history, then aggregates them according to a questionnaire-derived preference. In evaluation, preference settings aligned with corresponding held-out demonstrations, and participants reported lower workload scores with recommendations, though the fixed-order questionnaire makes that difference descriptive rather than causal.

  • Problem

    Traversability methods for off-road teleoperation often neglect operators’ differing route and speed preferences, despite those preferences affecting navigation choices.

    Source: §I · pp. 1–2, A. Problem Definition · p. 2

  • Method

    PATH-Net learns candidate cost and speed maps from demonstrations and uses a questionnaire-derived parameter for upper-tail cost aggregation and quantile-based speed selection.

    Source: §I · p. 2, A. Problem Definition · p. 3

  • Results

    The aggressive and conservative settings achieved the lowest pMHD and sMHD for their corresponding held-out demonstrations; mean NASA-TLX-based workload scores were 22.2 with recommendations versus 34.7 without.

    Source: §VI · p. 10

  • Takeaways & Limitations

    The results support further investigation of preference-conditioned trajectory and speed recommendations for teleoperated off-road navigation.

    Source: §VI · p. 10

    Evaluation was limited by participant numbers, including one held-out participant per preference category in validation and test sets, and by the fixed-order questionnaire design.

    Source: §VI · p. 10

Abstract

from arXiv · show

Traversability estimation is critical for ground vehicle teleoperation, yet existing methods often neglect the human operator's preferences for speed and risk tolerance. To address this gap, we propose PATH-Net, a Preference-Aware Traversability estimation Hierarchical Network that adaptively generates preference-conditioned traversability cost maps and speed profiles for off-road teleoperation. Our framework integrates visual and geometric bird's-eye-view features with short-term historical vehicle states through a multi-head shared-backbone architecture. Our PATH-Net infers traversability cost maps while optimizing speed recommendations by using maximum entropy inverse reinforcement learning and sparse speed supervision. Our method incorporates operator preference via an asymmetric feature aggregation mechanism, splitting inference into a cost branch that prioritizes high-risk traversability regions and a speed branch that outputs preference-adjusted recommended speeds. Experiments with 15 operators across 225 teleoperated trajectories show alignment with user-specific navigation strategies, achieving modified Hausdorff distance of 0.17 m (aggressive) and 0.15 m (conservative). In the questionnaire assessment, participants reported lower NASA-TLX workload scores with PATH-Net recommendations than without recommendations, decreasing from 34.7 to 22.2. These results support preference-conditioned trajectory and speed recommendation for off-road teleoperation.

I. INTRODUCTION

Off-road teleoperation can leave operators with delayed visual and coarse vehicle-state feedback, while existing traversability methods often neglect individual preferences and recent vehicle motion. PATH-Net is designed to generate cost and speed recommendations conditioned on both the terrain and a questionnaire-derived operator preference.

I. INTRODUCTION

  • NLOS teleoperation limits operators to delayed visual feeds and coarse vehicle-state telemetry, without vestibular or tactile cues.These sensory constraints are associated with high cognitive load, compromised depth perception, and degraded situational awareness.Source: §I · p. 1
  • Existing traversability methods are often operator-agnostic and static, despite operators choosing different routes and speeds for the same terrain.The resulting mismatch with risk-averse operators’ judgments can erode trust and trigger automation rejection; recent vehicle motion also provides context beyond a single terrain frame.Source: §I · p. 2
  • PATH-Net learns candidate cost functions and speed maps from operator demonstrations, then uses questionnaire-derived preference to aggregate them asymmetrically.One trained model can provide aggressive, neutral, or conservative advice without separate models for different operators.Source: §I · p. 2

A. Problem Definition

  • The problem formulation calls for both spatial guidance and speed recommendations, using terrain features alongside recent vehicle-state context.The local environment combines visual appearance and geometric terrain properties, while the history includes velocity, angular velocity, roll, and pitch.Source: A. Problem Definition · pp. 2–3
  • PATH-Net maps lower preference values to efficiency-oriented, higher-speed advice and higher values to cautious advice emphasizing higher-cost candidates and lower speeds.The preference parameter affects outputs through aggregation rather than being provided to the prediction network.Source: A. Problem Definition · p. 3

B. Framework Overview

PATH-Net combines visual-geometric BEV features with recent vehicle-state history to predict candidate cost and speed maps. Preference-dependent aggregation produces final maps that guide a path planner.

B. Framework Overview

  • Maximum Entropy IRL learns candidate cost maps to match operator preferences, while speed maps are learned from speed supervision.The prediction heads generate diverse candidate maps without explicit preference specialization.Source: B. Framework Overview · p. 3
  • At inference, upper-tail cost aggregation and quantile-based speed selection convert candidate maps into preference-dependent outputs for the path planner.The recent vehicle states and visual-geometric BEV features are fused before candidate prediction.Source: B. Framework Overview · pp. 3–4, Figure 3 · p. 4

III. THE PROPOSED PATH-NET

  • A dual-branch feature extractor captures environmental and vehicle-motion features before a shared-backbone, multi-head network predicts candidate cost and speed maps.The environmental branch fuses LiDAR geometry and camera features; a TCN encodes kinematic history.Source: B. Framework Overview · p. 3, §III · p. 4

A. Dual-Branch Feature Extraction

PATH-Net builds environmental features by fusing visual appearance with LiDAR geometry and encodes recent vehicle motion with a temporal convolutional network. The resulting branches support motion-aware traversability prediction on the BEV grid.

1) Environmental Perception and Fusion:

  • The environmental branch uses GFN to fuse geometric and visual BEV features, with visual features guiding how LiDAR geometry is transformed at each grid cell.The LiDAR guidance map contains height statistics, planarity, linearity, and surface-normal components; camera features add appearance cues for terrain with similar geometry but different surface properties.Source: A. Dual-Branch Feature Extraction · p. 4, 1) Environmental Perception and Fusion · p. 4

A. Dual-Branch Feature Extraction

2) Kinematic State Encoding:

  • A TCN encodes recent vehicle-state history with dilated causal convolutions, so each layer depends only on present and past inputs.Its output summarizes speed changes, angular-rate variations, and roll/pitch responses for subsequent cost and speed prediction, rather than predicting future vehicle states.Source: 2) Kinematic State Encoding · pp. 4–5

3) Feature Fusion:

The model tiles its kinematic embedding across the BEV grid and concatenates it with environmental features. A shared backbone and specialized heads then produce normalized candidate cost and speed maps.

3) Feature Fusion:

  • PATH-Net spatially tiles the kinematic embedding to match the environmental feature map, then concatenates the two for prediction at each BEV cell.This lets the shared backbone use local terrain features together with recent vehicle-motion context.Source: 3) Feature Fusion · p. 5

B. Multi-Head Candidate Cost and Speed Map Learning

  • A shared convolutional backbone and specialized output heads generate candidate cost and speed maps from the fused feature map.Heads are not assigned to particular preferences; independent initialization and preference-balanced bootstrap samples encourage diverse candidates, while preference labels are used only for sampling.Source: B. Multi-Head Candidate Cost and Speed… · p. 5

1) Output Parameterization and Normalization:

  • Cost logits are sigmoid-bounded as relative traversal costs, while speed predictions are normalized to physical units using the dataset maximum of 1.2 m/s.The speed normalization is preference-invariant; preference affects the final speed map through aggregation, and the map represents recommended speed limits rather than absolute safety boundaries.Source: 1) Output Parameterization and… · p. 5

2) Cost Map Learning:

PATH-Net learns traversability costs from expert and learner visitation frequencies using maximum-entropy IRL, while sparse trajectory speed labels supervise dense speed-map predictions.

2) Cost Map Learning:

  • Maximum-entropy IRL learns each cost map by comparing normalized expert and learner visitation frequencies over the BEV grid.The cost-map gradient reduces costs for demonstrator-visited states and increases costs for regions frequently visited by the learner.Source: 2) Cost Map Learning · p. 5
  • MPPI-based trajectory rollouts estimate learner visitation frequency by projecting vehicle poses onto grid cells and normalizing their importance weights.This formulation avoids explicit discretization of the vehicle control space.Source: 2) Cost Map Learning · p. 5

3) Speed Map Learning:

  • Sparse speed labels are dilated to nearby BEV cells within a 0.5 m radius, assigning each cell the nearest trajectory-point speed.The dilation does not affect collision checking.Source: 3) Speed Map Learning · p. 5
  • Despite supervision only in the labeled region, the fully convolutional network produces a dense speed map during inference.The training objective combines speed learning with the head-specific stratified-bootstrap mini-batch.Source: 3) Speed Map Learning · pp. 5–6

C. Preference Adaptation via Asymmetric Aggregation

PATH-Net conditions candidate cost and speed maps on preference through separate aggregation operators: an upper-tail mean for costs and a nearest-rank quantile for speeds.

C. Preference Adaptation via Asymmetric Aggregation

  • The preference parameter ρ maps linearly to aggregation weight α, with smaller α selecting higher speed candidates and larger α selecting lower speed candidates.The speed operator selects a single quantile rather than averaging the lower tail, avoiding accumulation of multiple low-speed predictions.Source: C. Preference Adaptation via… · p. 6
  • Cost maps use an upper-tail mean based on CVaR principles, while speed maps use nearest-rank quantile selection across candidate maps.The aggregation is applied grid cell-wise; α is an operational aggregation parameter rather than a true probability measure, and the implementation uses N = 5.Source: C. Preference Adaptation via… · p. 6
  • For N = 5, α = 0.95 selects the maximum candidate cost, whereas α = 0.05 averages all five candidates under the nearest-rank approximation.The cost aggregation is a finite-sample upper-tail mean.Source: C. Preference Adaptation via… · p. 6
  • Sorting cost and speed candidates independently at each grid cell may break the original pairing between each head’s cost and speed predictions.This is a consequence of the marginal aggregation procedure.Source: C. Preference Adaptation via… · p. 6

IV. EXPERIMENTAL SETUP AND DATA COLLECTION

The study assessed operator preference with a five-item questionnaire and collected teleoperation data from 15 trained participants, grouping them into aggressive, neutral, and conservative categories.

A. Participants and Preference Assessment

  • Fifteen participants completed the study, with an average of over 8 hours of teleoperation practice before data collection.The participants had a mean age of 24.0 years (SD = 2.4).Source: A. Participants and Preference… · p. 6
  • A five-item questionnaire based on established RTP measures assessed self-perceived control ability and efficiency-safety tradeoff preferences.Items used five-point Likert scales, and Q5 was inverted because it reflected safety-oriented behavior.Source: A. Participants and Preference… · p. 6
  • The total preference score R_u ranges from 5 to 25, with higher values indicating efficiency-oriented operation and lower values indicating conservative operation.Participants were classified using predetermined thresholds: aggressive at R_u ≥ 18, neutral at 14 ≤ R_u ≤ 17, and conservative at R_u ≤ 13.Source: A. Participants and Preference… · p. 6
  • The resulting groups comprised 4 aggressive, 6 neutral, and 5 conservative participants, each assigned a preference parameter ρ and aggregation level α.The mapping from ρ to α is defined in Section III-C.Source: A. Participants and Preference… · p. 6

B. Experimental Platform and Apparatus

The experiments used a sensor-equipped UGV on varied hillside terrain, with 15 operators completing repeated trials; sensor streams and demonstrations were converted into synchronized BEV inputs and speed supervision.

B. Experimental Platform and Apparatus

  • The UGV combined LiDAR geometry, camera imagery, and 10 Hz GNSS/RTK-GPS vehicle-state measurements for environmental and motion-history inputs.The camera captured 1280 × 720 RGB images, while the AHRS provided linear velocity, angular rate, and roll/pitch measurements.Source: B. Experimental Platform and Apparatus · pp. 6–7

C. Scenarios and Data Collection

  • Fifteen operators each completed 15 trials on hillside terrain, producing 225 trajectories with route choices spanning gentler and steeper slopes.The site included inclines up to 30°, ditches 0.1–0.2 m deep, and mixed vegetation; alternative routes included slopes ≤15° and 15°–30°.Source: C. Scenarios and Data Collection · p. 7

D. Data preprocessing

  • Sensor data was projected onto a 10 m × 10 m BEV grid at 0.1 m resolution, with LiDAR encoding elevation and occupancy and DINOv2-Small extracting visual features.Vehicle states were sampled at 10 Hz and segmented into overlapping 3.0 s windows; demonstrated speeds were Gaussian-smoothed with a five-frame standard deviation.Source: D. Data preprocessing · p. 7

V. RESULT ANALYSIS AND DISCUSSION

PATH-Net is evaluated for training behavior and alignment with held-out operator demonstrations, using participant-disjoint data and a shared MPPI planner. Its preference settings align with the corresponding held-out participants’ route and speed behavior, while validation pMHD is lower than for the adapted baselines.

1) Dataset Partition:

  • The participant-disjoint split assigns 9 participants and 135 trajectories to training, and 3 participants and 45 trajectories each to validation and testing.Each evaluation split has one participant per preference category, so preference-specific findings are held-out participant evaluations, not population estimates.Source: 1) Dataset Partition · p. 7

3) MPPI Planner:

  • MPPI estimates learner state-visitation frequencies for IRL and generates recommended trajectories, evaluating rollouts with traversal cost, speed-bound penalties, and terminal goal distance.Source: 3) MPPI Planner · p. 8

B. Training Convergence and Validation Alignment

  • The composite, IRL, and speed-regression losses decrease initially and plateau at approximately 20–30 training epochs.The reported convergence pattern is shared across the joint objectives.Source: B. Training Convergence and Validation… · p. 8
  • PATH-Net achieves lower validation pMHD than the adapted HDIF and Velociraptor baselines under identical 50-epoch training conditions.Validation pMHD is used because raw training losses are not directly comparable across methods with different objectives and loss scales.Source: B. Training Convergence and Validation… · p. 8

C. Offline Trajectory Alignment

  • Aggressive PATH-Net has the lowest mean pMHD and sMHD on held-out aggressive trajectories, while conservative PATH-Net has the lowest values on held-out conservative trajectories.Values increase when either setting is evaluated on the other preference group; standard deviations reflect trajectories from one held-out participant per group.Source: C. Offline Trajectory Alignment · pp. 8–9, Table II · p. 8
  • For a representative held-out case, PATH-Net produces localized high-cost regions while preserving a traversable corridor; HDIF’s speed visualization is a planner bound, not a learned speedmap.PATH-Net and Velociraptor produce spatial speedmaps.Source: Figure 6 · p. 8, C. Offline Trajectory Alignment · p. 9

D. Questionnaire-Based Workload Assessment

The paper assesses expected workload with and without PATH-Net recommendations and examines how preference settings change recommended operating points, maps, and routes. Participants reported lower workload with recommendations, while the preference examples show distinct speed, cost, and clearance trade-offs.

D. Questionnaire-Based Workload Assessment

  • The within-subject assessment compared expected workload without additional traversability information against workload after participants reviewed PATH-Net route and speed recommendations.Source: D. Questionnaire-Based Workload… · p. 9
  • NASA-TLX performance scores were reverse-scored so that higher values across all six subscales indicate greater reported workload.Source: D. Questionnaire-Based Workload… · p. 9
  • Participants reported a mean NASA-TLX overall workload score of 22.2 with PATH-Net recommendations versus 34.7 without them, a 36.0% relative decrease.Because the two conditions were presented in fixed order, the comparison is descriptive rather than causal evidence of workload reduction.Source: D. Questionnaire-Based Workload… · p. 9, Table III · p. 9

E. Preference Effects on Recommended Trajectories

VI. CONCLUSION AND FUTURE WORK

PATH-Net generates preference-conditioned cost and speed maps using asymmetric aggregation, with outputs varying by a questionnaire-derived preference parameter. Evaluation found preference-specific alignment and descriptive lower workload scores, while participant numbers and fixed questionnaire order limit the evidence.

VI. CONCLUSION AND FUTURE WORK

  • PATH-Net combines visual–geometric BEV perception and recent vehicle-motion history to generate candidate cost and speed maps, then aggregates them according to questionnaire-derived preference.Upper-tail aggregation is applied to candidate costs, while candidate speeds are selected by quantile.Source: §VI · p. 10
  • 22.2 versus 34.7 was the mean NASA-TLX-based workload score with versus without PATH-Net recommendations, a descriptive difference given fixed presentation order.Aggressive and conservative settings achieved the lowest pMHD and sMHD values for their corresponding held-out demonstrations; conservative recommendations had lower speed and path-averaged cost, with larger obstacle clearance.Source: §VI · p. 10
  • The evaluation was limited by participant numbers, including one held-out participant per preference category in validation and test, and by fixed-order questionnaire design.Future work will examine online preference estimation, additional vehicle–terrain measurements, larger participant samples, and counterbalanced protocols.Source: §VI · p. 10
Loading 2610.30605v1…