PaperlayerView exact revision on arXiv

Source-linked AI summary

T-ConSG: A LiDAR-Based Real-Time Temporal Construction Scene Graph for Dynamic and Evolving Environments

Yilong Chen, Yong K. Cho

arXiv:2610.34041v1cs.RO

TL;DR

Dynamic and long-term SLAM methods detect motion and environmental change, but their metric outputs offer limited structure for higher-level spatial and temporal reasoning. T-ConSG incrementally builds a LiDAR-based scene graph of free space, objects, and environmental events. Across navigation, progress monitoring, and change-detection studies, it demonstrates a unified representation for reasoning about transient dynamics and persistent changes.

  • Problem

    Dynamic and long-term SLAM methods detect motion and environmental changes, but their predominantly metric outputs provide limited structure for higher-level spatial and temporal reasoning.

    Source: §I · p. 1

  • Method

    T-ConSG incrementally constructs a LiDAR-based temporal scene graph jointly representing free space, static and dynamic objects, and environmental events.

    Source: §I · pp. 1–2

  • Results

    Across navigation, progress monitoring, and online change detection, T-ConSG demonstrates a unified representation for organizing spatial and temporal information and supporting reasoning about transient dynamics and persistent changes.

    Source: A. Dynamics-Aware Navigation · p. 5, B. Construction Progress Monitoring · p. 6, C. Online Change Detection · p. 7

  • Takeaways & Limitations

    The three application studies show T-ConSG's applicability to construction environments containing short-term dynamics and persistent changes.

    Source: §I · p. 2, §V · p. 8

    The current implementation uses 2D spatial representations; extension to real-world 3D environments remains future work.

    Source: §V · p. 8

Abstract

from arXiv · show

Dynamic and continuously evolving environments present significant challenges for robotic mapping and scene understanding, particularly on construction sites where moving agents, temporary obstacles, and persistent structural changes coexist. While dynamic and long-term SLAM methods can detect motion and environmental changes, their outputs are predominantly metric representations that provide limited structure for higher-level spatial and temporal reasoning. This paper presents a LiDAR-based real-time temporal construction scene graph (T-ConSG) that jointly represents free-space regions, static objects, dynamic objects, and environmental events with spatial-relational and event-participation edges. The proposed framework incrementally constructs the graph from LiDAR-derived geometric evidence, maintaining persistent free-space and static-object information while tracking dynamic objects. Changes in accumulated spatial evidence are further represented as explicit installation and removal events, allowing the graph to capture both transient dynamics and persistent environmental evolution. We demonstrate the proposed representation through three complementary application studies: dynamics-aware navigation in a bricklaying testbed, construction progress monitoring on the ConSLAM dataset, and online change detection in a simulated construction environment. The results demonstrate that the proposed T-ConSG provides a unified representation for organizing spatial and temporal information and supports robotic reasoning across environments containing both short-term dynamics and persistent changes.

I. INTRODUCTION

Construction sites change continuously, so robots need spatial knowledge that distinguishes transient motion from persistent environmental change. T-ConSG addresses gaps in metric and temporal scene-graph representations by structuring free space, objects, and environmental events.

I. INTRODUCTION

  • Construction sites combine moving agents, temporary obstacles, and persistent structural changes, requiring spatial knowledge that preserves how the workspace evolves.Robots must distinguish transient motion from lasting environmental change, not only estimate current geometry.Source: §I · p. 1
  • Existing dynamic and long-term SLAM methods primarily maintain localization, metric reconstruction, object states, or geometric change estimates, rather than structuring evidence for higher-level reasoning.The paper identifies entities and explicit events as a way to organize such evidence for spatial and temporal reasoning.Source: §I · p. 1

A. 3D and 4D Scene Graphs

  • Temporal scene graphs remain comparatively immature in representing persistent changes to spatial entities and relationships, while free space is often treated mainly as navigation support.Prior work on 3D and 4D scene graphs offers hierarchical representations, but temporal interaction between free space, objects, and changes has received less attention.Source: §I · p. 1, A. 3D and 4D Scene Graphs · p. 2

I. INTRODUCTION

  • T-ConSG jointly represents free-space regions, static structures, tracked dynamic objects, and explicit temporal events with spatial-relational and event-participation edges.It incrementally maintains spatial state from LiDAR evidence and distinguishes transient motion from persistent structural changes.Source: §I · p. 1
  • The paper presents a real-time incremental LiDAR mapping framework and demonstrates it in dynamics-aware navigation, construction progress monitoring, and online change detection.These applications illustrate the representation's use in construction-oriented robotics.Source: §I · p. 2

B. Dynamic and Long-Term SLAM

Dynamic and long-term SLAM address motion and environmental change, but construction settings require representations that also organize persistent spatial state and temporal events. T-ConSG builds such a graph from LiDAR mapping outputs, using a planar implementation for ground-robot navigation and change detection.

B. Dynamic and Long-Term SLAM

  • Dynamic SLAM addresses moving entities, while long-term methods address changes revealed when mapped areas are revisited; both still face difficult association and change-detection problems.The review distinguishes transient motion from changes that occur outside the sensor's view and appear abrupt on revisit.Source: B. Dynamic and Long-Term SLAM · p. 2

III. METHODOLOGY

  • The graph is constructed from Dynablox mapping outputs, with DLIO used as the default odometry backend for the quadruped platform.The listed outputs include ever-free, TSDF, and never-free maps, along with static and dynamic objects.Source: §III · p. 2
  • T-ConSG represents free space, static objects, tracked dynamic objects, and environmental events as vertex families linked by spatial and temporal-participation edges.Event types include installation, removal, appearance, disappearance, and observation.Source: §III · p. 2
  • The current implementation projects 3D LiDAR and volumetric-map observations onto a horizontal plane for ground-robot navigation and structural-change detection.The node, relation, and event definitions can generalize to 3D voxels, but efficient volumetric connectivity and boundary extraction remain future implementation work.Source: §III · pp. 2–3

A. Free-Space Node

The free-space node persistently maps historically observed free space from LiDAR evidence. Its connected components and polygon boundaries support spatial-overlap and temporal-change reasoning, with confirmed installations removed from the accumulated map.

A. Free-Space Node

  • The free-space node accumulates historically observed free cells, incrementally merges four-connected components, and traces polygon boundaries for spatial-overlap and change reasoning.Inner boundaries mark areas excluded from confirmed historical free space; they are not interpreted as obstacles without additional evidence.Source: A. Free-Space Node · p. 3
  • Cells associated with a confirmed installation event are removed from the accumulated free-space representation, after which affected connectivity and polygon geometry are rebuilt.This lets the map reflect newly occupied regions.Source: A. Free-Space Node · p. 3

B. Dynamic-Object Node

The dynamic-object node turns frame-level detections into temporally consistent tracked states. The static-object node accumulates persistent cells from map evidence while excluding dynamic detections and allowing explicit removal events to update the map.

B. Dynamic-Object Node

  • An EKF associates Dynablox detections with active tracks using object identity and spatial and temporal thresholds, then propagates and corrects states to estimate continuous trajectories and velocity.When association criteria fail, the previous track ends and a new EKF instance is initialized; the thresholds are empirically fixed.Source: B. Dynamic-Object Node · p. 3

C. Static-Object Node

  • Static-object cells require both never-free map evidence and a static object-level observation, with dilated dynamic detections removed before persistent accumulation.The static map retains previously detected cells during normal observation, while explicit removal events can delete cells.Source: C. Static-Object Node · pp. 3–4

D. Event Node

Event nodes represent persistent structural changes by identifying conflicts between accumulated free-space and static-object maps. Candidate changes are stabilized over time before being published as installation or removal events.

D. Event Node

  • Overlapping accumulated free-space and static-object cells form potential change regions, with TSDF support distinguishing installation candidates from removal candidates.A signed temporal counter and a persistence threshold suppress transient inconsistencies; confirmed cells must also remain spatially unchanged for consecutive updates before publication.Source: D. Event Node · p. 4
  • Confirmed candidate cells are grouped by four-neighbor connectivity and converted into event polygons that can update the affected free-space or static-object representation.Each event node records the event type, affected polygonal region, and grid-cell indices.Source: D. Event Node · p. 4
  • Event nodes can also record dynamic-object appearances and disappearances or the first observation of a static object.Source: D. Event Node · p. 4

E. Graph Edge Construction

Graph edges separate spatial relationships among world-state nodes from event-participation links that associate temporal changes with affected entities. This represents both the current scene configuration and its evolution.

E. Graph Edge Construction

  • Spatial-relational edges connect world-state nodes to express spatial or accessibility relationships.For example, dynamic objects inside a free-space region can access static objects connected to that same region.Source: E. Graph Edge Construction · p. 4
  • Dynamic objects inside a free-space node and static objects connected to it are linked by directional access relations.The edge labels express the two directions of the same access relation.Source: E. Graph Edge Construction · p. 4
  • Event-participation edges connect event nodes with world-state nodes and use labels to identify event types, not opposite directions of one relation.Source: E. Graph Edge Construction · p. 4
  • Event relations distinguish dynamic-object appearance or disappearance, static-object observation or removal, and free-space observation or installation.Source: E. Graph Edge Construction · p. 5
  • Spatial-relational edges describe the current configuration, while event-participation edges record how it evolves over time.Source: E. Graph Edge Construction · p. 5

IV. APPLICATION AND DISCUSSION

The paper evaluates T-ConSG across complementary construction-related datasets and applications; it ran at a real-time processing rate on the robot’s onboard CPU.

IV. APPLICATION AND DISCUSSION

  • Three complementary construction-related datasets were used because no single available dataset covered all T-ConSG capabilities.Source: §IV · p. 5
  • The demonstrations cover navigation, construction progress monitoring, and online change detection.Source: §IV · p. 5
  • T-ConSG ran at a real-time processing rate on the robot’s onboard CPU in all experiments.Source: §IV · p. 5

A. Dynamics-Aware Navigation

In the bricklaying testbed, the robot navigated around a human participant using a social-force model informed by T-ConSG dynamic-object nodes. The adjusted trajectories increased minimum time-to-collision while maintaining similar closest pedestrian distance.

A. Dynamics-Aware Navigation

  • The robot and a human built brick walls concurrently, with waypoint encounters designed to include merging, passing, and crossing.Source: A. Dynamics-Aware Navigation · p. 5
  • The social-force model applies a repulsive force from human coworkers as a displacement to planner-generated waypoints, using T-ConSG dynamic-object nodes.The evaluation used path length, path irregularity, closest pedestrian distance, and minimum time-to-collision across 10 trials.Source: A. Dynamics-Aware Navigation · p. 5
  • Mean minimum time-to-collision rose from 4.76 s to 5.11 s, while mean closest pedestrian distance was similar at 2.62 m versus 2.60 m.The comparison is reported for SFM-adjusted trajectories; Table I reports means ± standard deviations across 10 trials.Source: A. Dynamics-Aware Navigation · p. 5, Table I · p. 5
  • The reported adjustment traded path efficiency and smoothness for modest timing changes, mainly changing avoidance timing and shape rather than ultimate spatial clearance.Source: A. Dynamics-Aware Navigation · p. 5

B. Construction Progress Monitoring

Across ConSLAM Sequences 2–5, T-ConSG recovers most ground-truth free space and maps increasing static-object area as construction progresses. Inter-sequence comparisons also identify installed and removed regions.

C. Online Change Detection

In five simulated trials, T-ConSG detected structural changes after the robot revisited affected locations that had changed outside its field of view. The graph distinguished these persistent changes from moving workers and updated inferred accessibility.

  • All 10 structural changes across five trials were correctly classified without false event detections.The simulated site included one installation and one removal per trial, both occurring outside the robot’s field of view.Source: Figure 7 · p. 7, C. Online Change Detection · p. 7
  • Revisiting a changed location lets current TSDF evidence classify conflicts with persistent free-space or static-object maps as installations or removals.Moving workers were represented as dynamic objects and did not produce persistent structural-change events.Source: C. Online Change Detection · p. 7
  • After a structural change disconnects the free-space and static-object nodes, the graph removes the corresponding inferred accessibility relation.The trial queried access from point A in a free-space node to point B in a static-object node.Source: C. Online Change Detection · pp. 7–8

V. CONCLUSION AND FUTURE WORK

T-ConSG integrates persistent spatial evidence with explicit environmental events to represent dynamic and evolving environments. The work demonstrates its applicability while identifying limitations and future extensions.

  • T-ConSG integrates persistent spatial evidence and explicit environmental events to support reasoning about transient dynamics and persistent changes.Source: §V · p. 8
  • The current implementation uses 2D spatial representations; future work will extend it to real-world 3D environments.Source: §V · p. 8
  • Future work will more explicitly use free-space regions for accessibility reasoning and temporal change querying, and incorporate semantics to associate geometric changes with construction elements and activities.Source: §V · p. 8

Related papers

Loading 2610.34041v1…