PaperlayerView exact revision on arXiv

Source-linked AI summary

Experience-Driven Continual Learning of Terrain Traversability for Quadruped Robots

Luca Bricarello, João Carlos Virgolino Soares, Alberto Sanchez-Delgado, Fulvio Mastrogiovanni, Claudio Semini

arXiv:2609.39755v1cs.RO

TL;DR

Quadruped navigation needs to anticipate terrain-robot interaction because appearance and geometry alone do not reveal contact outcomes, and continual learning can forget prior terrain encounters. The paper predicts confidence-weighted contact indicators from pre-contact images and gates replay-based updates using recent and historical validation; in a sequential hardware stream, this reduced final anchor NLL degradation by 23.1% with similar new-terrain adaptation.

  • Problem

    Terrain geometry and appearance alone do not reveal how a quadruped will slip, load its feet, or expend energy, while online finetuning may forget earlier terrain encounters.

    Source: §I · p. 1

  • Method

    The pipeline predicts five confidence-weighted foot-contact indicators from pre-contact visual features and uses replay with recent and historical held-out validation to gate model updates.

    Source: §I · pp. 1–2

  • Results

    The historical validation gate reduced final anchor NLL degradation by 23.1% relative to replay with recent validation alone, with similar new-terrain adaptation in a one-stream hardware ablation.

    Source: §I · p. 2

  • Takeaways & Limitations

    The results support guarded adaptation under the tested conditions.

    Source: §V · p. 8

    The results do not establish an overall accuracy advantage or isolate the gate’s effect on navigation.

    Source: §V · p. 8

Abstract

from arXiv · show

Safe and efficient quadruped navigation over unfamiliar terrain requires predicting terrain-robot interaction before contact: geometry and visual appearance alone cannot reveal how the robot will slip, load its feet, or expend energy. This paper presents a continual learning pipeline that uses locomotion experience to learn these interaction outcomes from pre-contact images and continually updates the predictions as new contacts are observed. Pre-contact descriptors, produced by a DINOv3 backbone model frozen during training, are mapped to five proprioceptive indicators weighted according to measurement reliability: planar foot slip, mean normal ground-reaction force, traction index, cost of transport, and touchdown loading rate. A compact evidential regressor allows us to predict these indicators together with aleatoric and epistemic uncertainty from the visual descriptors. Continual adaptation combines bounded experience replay with a validation gate: candidate models replace the deployed predictor only when they improve performance on recent held-out data while keeping degradation on historical held-out data within a prescribed tolerance. Predictions and epistemic uncertainty are projected into a local multilayer map and combined into a conservative traversability score map whose property weights can be adjusted without retraining. The resulting map is used for downstream navigation tests. The ROS2 implementation supports evaluation on a Unitree Go2 in simulation and on hardware, with models trained separately in each domain. On a sequential hardware stream over three previously unseen terrains, gated replay reduces final anchor negative log-likelihood (NLL) degradation by 23.1% relative to replay without the gate while attaining similar new-terrain adaptation.

I. INTRODUCTION

Terrain appearance and geometry do not fully reveal how a quadruped will interact with terrain, while continual adaptation risks forgetting earlier encounters. The paper predicts multiple contact outcomes from pre-contact vision and uses replay with held-out validation to gate model updates.

I. INTRODUCTION

  • Terrain elevation maps support foothold and path planning, but identical geometry can conceal differences in friction, compliance, or load-bearing properties,,,.Vegetation that appears to block a path may instead be pushed through,.Source: §I · p. 1
  • Online finetuning may forget prior terrain encounters, and models may become overconfident on out-of-distribution observations, [13],,.Visual and proprioceptive sensing offer complementary information about terrain before contact and its response during locomotion,.Source: §I · p. 1
  • The system maps pre-contact visual descriptions to five confidence-weighted foot-contact indicators and predicts each indicator with uncertainty.The indicators are planar foot slip, mean normal ground-reaction force, traction index, cost of transport, and touchdown loading rate.Source: §I · pp. 1–2
  • The historical validation gate reduced final anchor NLL degradation by 23.1% relative to replay with recent validation alone, with similar new-terrain adaptation.This is a one-stream hardware ablation.Source: §I · p. 2
  • The property-resolved evidential map retains individual interaction predictions and epistemic uncertainties, allowing navigation preferences and uncertainty penalties to change without retraining.Source: §I · p. 2

II. RELATED WORK

Prior work addresses geometric and semantic traversability, proprioceptive assessment, and learning from experience, but these approaches leave gaps in linking a robot’s contact outcomes to pre-contact vision and in deciding whether continual updates should be deployed.

A. Geometric and Semantic Traversability

  • Geometric maps and semantic annotations support planning, but do not directly measure the target robot’s foot slip, normal force, or cost of transport.Geometric maps may also miss the support surface beneath vegetation.Source: A. Geometric and Semantic Traversability · p. 2

B. Proprioceptive Assessment and Physical Properties

  • The formulation links five interaction indicators from completed stances to pre-contact visual observations, weighting targets by their individual measurement confidence.The stated gap is connecting multiple effects measured during the robot’s own contacts to what it saw beforehand while accounting for measurement reliability.Source: B. Proprioceptive Assessment and… · p. 2

C. Learning from Experience and Continual Adaptation

  • Unlike prior retention approaches based on selecting, replaying, or generating past experience, this approach separately gates model replacement using recent and historical held-out data.Acceptance requires improved predictive likelihood on recent experience while limiting historical degradation relative to the current model.Source: C. Learning from Experience and… · pp. 2–3

D. Uncertainty and the Navigation Interface

The system distinguishes unfamiliarity from poor traversability and retains separate property and epistemic-uncertainty layers before aggregating them into a navigation cost. It defines traversability through robot-terrain interaction outcomes rather than semantic terrain categories.

D. Uncertainty and the Navigation Interface

  • Unfamiliarity and poor traversability are distinct, and uncertainty-aware methods represent terrain properties or predictive uncertainty for navigation,,,,.The paper positions its representation as predicting individual interaction outcomes and their uncertainties.Source: D. Uncertainty and the Navigation… · p. 3

A. Problem Formulation and System Overview

  • Traversability is framed as expected robot-terrain interaction outcomes, represented by five indicators that are converted to property-specific scores and fused into a navigation cost.The definition is robot-dependent rather than a semantic terrain category.Source: A. Problem Formulation and System… · p. 3
  • A frozen DINOv3 backbone extracts spatial visual features, which are associated with completed foot-contact targets and confidence values.The five targets are planar slip, normal ground-reaction force, traction, cost of transport, and touchdown loading rate.Source: A. Problem Formulation and System… · p. 3
  • Training samples use visual observations acquired before contact, and deployment projects predictions into a local multilayer traversability map.Samples may be stored for offline training or used immediately by the continual learner.Source: A. Problem Formulation and System… · p. 3

B. Causal Visual–Proprioceptive Supervision

The system pairs pre-contact visual features with proprioceptive outcomes from the robot’s subsequent foot contact. Contact measurements are confidence-weighted, and the association uses only terrain visible before touchdown.

B. Causal Visual–Proprioceptive Supervision

  • A frozen DINOv3 backbone extracts local visual tokens, keeping descriptors compatible across offline and online updates to the downstream model.Images are resized to 576 × 576 pixels and represented as a 36 × 36 grid of local tokens.Source: B. Causal Visual–Proprioceptive… · p. 3
  • The supervision includes planar foot slip and force-based indicators: mean normal ground-reaction force and a traction index derived from tangential force relative to normal load.The traction index is not a direct estimate of terrain friction; planar slip separately penalizes loss of grip.Source: B. Causal Visual–Proprioceptive… · pp. 3–4
  • Cost of transport and touchdown loading rate complete the five contact indicators associated with pre-contact visual observations.The touchdown loading rate uses force samples from the 50 ms touchdown window.Source: B. Causal Visual–Proprioceptive… · pp. 4–5
  • Individual measurement confidences weight the learning objectives, while missing measurements receive zero confidence and contacts below the minimum-confidence threshold are rejected.A contact is rejected when its minimum property confidence is below 0.25.Source: B. Causal Visual–Proprioceptive… · pp. 4–5
  • Contact outcomes are matched to a feature map captured before touchdown, at most 5 s earlier, preventing future observations from leaking into training samples.The projected patch must cover the same 5 cm contact-association cell.Source: B. Causal Visual–Proprioceptive… · p. 5

C. Uncertainty-Aware Multi-Property Regression

A compact evidential regressor predicts the five interaction properties and their uncertainties. Its confidence-weighted training objective and property-specific normalization support uncertainty-aware, configurable traversability scoring.

C. Uncertainty-Aware Multi-Property Regression

  • A shared MLP trunk feeds five independent heads that output Normal–Inverse-Gamma parameters for the interaction properties.The predictive mean is γ, and the regressor produces the five properties and their uncertainties in one forward pass.Source: C. Uncertainty-Aware Multi-Property… · p. 5
  • Only epistemic uncertainty is propagated to the traversability map; aleatoric uncertainty is used for model evaluation.Source: C. Uncertainty-Aware Multi-Property… · p. 5
  • The confidence-weighted evidential objective combines Student-t negative log-likelihood with a penalty against confident but inaccurate predictions.Input and target scalers are fitted on the offline training partition and then frozen; timestamp- or descriptor-matched samples are grouped before splitting.Source: C. Uncertainty-Aware Multi-Property… · p. 5
  • Slip and cost of transport are sign-inverted before scoring, while the other three indicators retain their signs; property means are normalized using fixed reference percentiles.Variance is scaled by the squared percentile interval but not clipped, preserving large uncertainty on out-of-distribution terrain.Source: C. Uncertainty-Aware Multi-Property… · p. 5

D. Guarded Continual Adaptation

Continual adaptation trains candidates from recent experience augmented with bounded replay, then uses held-out recent and historical data to decide whether to promote them. The historical anchor remains separate from replay training.

D. Guarded Continual Adaptation

  • Recent samples are split into training and validation buffers, and near-identical descriptors are rejected before an update is initiated.A nominal update starts after 50 recent training samples and at least 10 held-out samples.Source: D. Guarded Continual Adaptation · p. 5
  • Each recent training block is combined with bounded replay, including an online memory capped at 1000 samples and sampled across planar-slip quantiles.Replay is used for gradient updates, while a separate anchor set stays held out to measure retention.Source: D. Guarded Continual Adaptation · p. 5
  • A candidate is promoted only if recent validation improves and historical anchor NLL degradation stays within tolerance; in the hardware experiment, the allowed per-promotion increase was 0.15%.Candidates with non-finite outputs are rejected, and the previous model continues inference when promotion fails.Source: D. Guarded Continual Adaptation · p. 5
  • Continual adaptation starts from an offline-trained checkpoint, with an initially held-out partition serving as the historical anchor.Source: D. Guarded Continual Adaptation · p. 5

E. Spatial Traversability Mapping

Predicted property means and epistemic variances are projected into a local grid map and fused into a configurable traversability score. Separate property and uncertainty layers are retained for navigation choices.

E. Spatial Traversability Mapping

  • Visual patch predictions are projected using synchronized depth and camera pose into a robot-centered grid map with 0.1 m resolution.Multiple patches overlapping a cell have their first and second moments aggregated.Source: E. Spatial Traversability Mapping · pp. 5–6
  • Before property fusion, each cell score is conservatively reduced by 0.15 times the square root of its variance, which includes epistemic uncertainty and overlap disagreement.Source: E. Spatial Traversability Mapping · p. 6
  • The weighted traversability score uses configurable property weights, while individual property and uncertainty layers remain available; unobserved cells stay unknown and severe low scores can be marked lethal.Changing the navigation objective does not require retraining the predictor.Source: E. Spatial Traversability Mapping · p. 6

A. Experimental Setup and Protocol

The evaluation separates simulation and hardware datasets and trains domain-specific models, using contact-derived supervision rather than surface identities as learning labels. Simulation evaluates two interaction indicators, while hardware evaluates all five.

A. Experimental Setup and Protocol

  • Simulation uses three visually distinct surfaces, while hardware property prediction and mapping use five surfaces and a separate three-terrain continual-learning stream.The hardware stream sequences artificial grass, textured foam, and foam; surface identities define evaluation, not learning labels.Source: A. Experimental Setup and Protocol · p. 6
  • Simulation scoring fuses planar foot slip and traction index; hardware scoring uses all five indicators because simulated rigid contact does not represent the broader interaction vector.The weights are wsim = [0.5, 0, 0.5, 0, 0] and wreal = [0.4, 0.05, 0.1, 0.05, 0.3].Source: A. Experimental Setup and Protocol · p. 6
  • Offline prediction evaluates held-out contact-level tests using a separate multi-terrain acquisition dataset per domain, split 70% training, 15% validation, and 15% testing, with contacts sharing timestamps or identical descriptors grouped.Source: Table II · p. 6
  • Simulation and hardware datasets are disjoint, and their regressors are trained independently without assuming transfer between domains.Offline prediction and closed-loop navigation are evaluated separately from the sequential hardware continual-learning comparison.Source: A. Experimental Setup and Protocol · p. 6

B. Property Prediction and Calibration

Offline prediction and mapping assess contact-indicator agreement and spatial separation, while a sequential hardware comparison tests whether historical validation gating improves retention beyond replay. The gate reduces final anchor NLL degradation while maintaining similar mean new-terrain adaptation.

B. Property Prediction and Calibration

  • Offline regressors use multi-terrain acquisition datasets and held-out contact-level visual-contact pairs; MAE, R2, and C80 assess agreement with proprioceptive indicators.Coverage is for the central 80% Student’s t predictive interval, incorporating aleatoric and epistemic uncertainty.Source: B. Property Prediction and Calibration · p. 6
  • Mapped-cell score gaps are 0.2991 in simulation and 0.4382 on hardware; the smaller simulated gap may reflect the camera’s distant terrain view.The comparison excludes mixed-support boundaries and averages within each surface and run.Source: B. Property Prediction and Calibration · p. 6

C. Continual Acquisition and Retention

  • 0.608% vs 0.790% final anchor NLL degradation: CL-P reduces degradation 23.1% relative to Replay, with mean new-terrain adaptation of 1.658% vs 1.664%.Table III averages five optimization seeds on one fixed stream; anchor NLL is measured on the gate-validation set.Source: Table III · p. 7, C. Continual Acquisition and Retention · p. 7
  • The sequential ablation starts both methods from a Lab-only checkpoint and replays the same recorded stream, with independent test acquisitions excluded from training, replay, validation, and promotion.The stream contains 951 accepted contacts across artificial grass, textured foam, and foam; new-terrain adaptation averages percentage MAE reductions across the three phases.Source: C. Continual Acquisition and Retention · p. 7
  • CL-P adapts more slowly on artificial grass (0.756% vs 1.890%), but gains on textured foam (3.239% vs 2.329%) and foam (0.979% vs 0.773%) compensate.Both methods continue improving on previously encountered online terrains in later phases, with no positive error increase after their respective phases in this short stream.Source: C. Continual Acquisition and Retention · p. 7
  • Across 80 candidate opportunities, CL-P promotes 53 and rejects 24 by the anchor gate, while Replay promotes 70; Fig. 3 tracks relative anchor NLL across terrain phases.The figure marks seed means and ranges, rejected CL-P candidates, and phase transitions.Source: C. Continual Acquisition and Retention · p. 7, Figure 3 · p. 7

D. Traversability-Aware Navigation

Simulation navigation tests whether traversability-aware mapping supports closed-loop planning over surfaces with different friction. Both Offline and CL-P completed all trials, while the conclusion limits the evidence to the tested conditions and notes the scope of the hardware comparison.

D. Traversability-Aware Navigation

  • Trials used Nav2 with the SMAC Lattice planner in simulation, with a goal 15 m away and a 180 s limit; success required reaching the goal without falling or operator intervention.The environment includes elevation changes and three simulated surfaces with different friction coefficients.Source: D. Traversability-Aware Navigation · pp. 7–8
  • Both methods completed all five navigation trials; mean unfavorable-terrain exposure was 3.13% for Offline and 2.76% for CL-P, with path lengths of 23.75 m and 24.53 m.The comparable results support preserved downstream navigation performance under these conditions, but do not isolate the gate’s contribution.Source: D. Traversability-Aware Navigation · p. 8

V. CONCLUSIONS

  • The pipeline predicts five interaction indicators, adapts through validation-gated replay, and maps property predictions and epistemic uncertainty for navigation.The reported results support guarded adaptation under tested conditions, but establish neither an overall accuracy advantage nor a gate effect on navigation.Source: §V · p. 8
  • The hardware comparison uses one recorded stream with five optimization seeds, and simulation and hardware models are trained separately.Future work will test longer terrain sequences, deformable terrain, varying locomotion conditions, uncertainty estimation, and active sampling on unfamiliar terrain.Source: §V · p. 8

Related papers

Loading 2609.39755v1…