PaperlayerView exact revision on arXiv

Source-linked AI summary

Enhancing Robotic Perception and Adaptability through Sensor Fusion and Origami-Inspired Designs

Namai Chandra, Jaison Jose, Kavi Arya, Shivaram Kalyanakrishnan

arXiv:2610.09828v1cs.RO

TL;DR

Compact robots must recover scene geometry despite changing lighting and surface texture while meeting tight payload and cost constraints. This prototype combines RGB-D and LiDAR depth fusion with origami-inspired wheels that sweep the scanner through changing elevations, with IMU alignment. In indoor and outdoor tests, invalid full-frame depth fractions fell from 21% to 11% and from 48% to 18%, respectively.

  • Problem

    Compact robots need usable 3D perception under changing lighting and surface texture while remaining small, lightweight, and inexpensive.

    Source: §I · p. 1

  • Method

    The prototype fuses LiDAR returns into RGB-D depth for RTAB-Map and reuses origami-inspired wheel motion to sweep the LiDAR through intermediate elevations, with IMU scan alignment.

    Source: §VIII · p. 9

  • Results

    Invalid full-frame depth fractions fell from 21% to 11% indoors and from 48% to 18% outdoors.

    Source: §VIII · p. 9

  • Takeaways & Limitations

    The wheel mechanism redirects the scanner using hardware already used for locomotion, and the present recordings do not quantify map-density gains from a complete transition.

    Source: §VII · p. 9

    Measurements cover one prototype, one indoor corridor, and one outdoor location, with three manually piloted runs per sensor configuration, limiting generalization beyond the tested settings.

    Source: A. Limitations · p. 9

Abstract

from arXiv · show

Compact mobile robots must recover scene geometry under changing lighting and surface texture while working within tight payload and cost limits. We present a compact mobile robot that uses origami-inspired wheels for locomotion and active control of its sensing geometry. As the wheels move between terrain-adaptive configurations, the changing chassis pitch sweeps a 2D LiDAR through intermediate elevations; held wheel positions provide a chosen viewing angle. An IMU accounts for chassis attitude, and a fusion node projects LiDAR returns into the RGB-D depth stream supplied to RTAB-Map. The arrangement uses the wheel actuation already present on a sub-300 USD, sub-2 kg prototype to extend the scanner's viewing geometry. We assess depth fusion in a textureless indoor corridor and an outdoor sunlit area, with three runs per sensor configuration in each setting. Mean full-frame invalid-depth fractions fell from 21% to 11% indoors and from 48% to 18% outdoors. The prototype combines improved depth coverage with a continuously adjustable LiDAR viewpoint using the same actuation that reconfigures its wheels.

I. INTRODUCTION

Compact robots face a trade-off between usable 3D perception and tight size, weight, and cost constraints. This work combines RGB-D and LiDAR sensing with origami-inspired wheels that also steer the LiDAR scan plane.

I. INTRODUCTION

  • Compact robots must provide usable 3D perception while remaining small, lightweight, and inexpensive for constrained search-and-rescue settings.The system addresses this challenge with an origami-inspired chassis whose wheel poses also steer a 2D LiDAR scan plane.Source: §I · p. 1

A. Depth Sensing on a Small Robot

  • Stereo depth can fail on textureless surfaces, while 2D LiDAR provides a single horizontal plane without visual information and 3D LiDAR is costly and heavy.The D435i also loses depth completeness with range and direct sunlight.Source: A. Depth Sensing on a Small Robot · p. 1

B. How Fusion and Morphing Are Combined

C. Scope and Positioning

The prototype brings established RGB-D–LiDAR fusion, 2D LiDAR scanning, and morphing-wheel techniques together on one compact robot. Its wheel motion changes the scanner viewpoint, while an IMU aligns scans and a fusion node supplies depth to a standard SLAM backend.

C. Scope and Positioning

  • The design connects established sensing and mechanical techniques on one compact robot, measures depth completeness under two sensor stressors, and documents wheel-driven scan motion.The design combines wheel-driven viewpoint changes, IMU scan alignment, and fused depth supplied to a standard SLAM backend.Source: C. Scope and Positioning · p. 2, Table I · p. 2

A. RGB-D and LiDAR Fusion

  • Prior work combines RGB-D and LiDAR for tracking, localization, mapping, or navigation; this work uses lightweight geometric fusion to supply RTAB-Map with a more complete depth image.The stated goal is to improve the depth image supplied to RTAB-Map on the Jetson Orin Nano.Source: A. RGB-D and LiDAR Fusion · p. 2

B. Active Scanning and Mapping

  • Dedicated actuators have been used to sweep 2D LiDAR; here, wheel motion changes scan-plane elevation but cannot position the scanner independently of the chassis.The design joins RGB-D–LiDAR integration to chassis-driven scanning.Source: B. Active Scanning and Mapping · p. 3

C. Origami-Inspired Wheels

  • The flasher-pattern-inspired wheels serve both terrain adaptation and viewpoint change, using wheel reconfiguration to enact active perception.Mechanical scanning is established; this prototype uses the wheel motion to supply scan motion on the same robot.Source: C. Origami-Inspired Wheels · p. 3

A. Coordinate Frames and Projection

  • The method uses an SE(3) transformation between LiDAR and camera frames, then projects camera-frame 3D points onto the image plane using the pinhole model.The transformation represents sensor alignment; the projection uses factory-calibrated camera intrinsics.Source: A. Coordinate Frames and Projection · p. 3

IV. METHODOLOGY

The perception pipeline preprocesses sensor data before it reaches RTAB-Map. It synchronizes and transforms LiDAR measurements, projects them into the camera image, and confidence-weights their fusion with RGB-D depth.

IV. METHODOLOGY

  • The pipeline routes raw sensor data through fusion and tilt compensation before the SLAM backend constructs a map.The perception module projects LiDAR returns into the RGB-D camera frame and fuses depth in the overlapping field of view before sending the stream to RTAB-Map.Source: §IV · p. 3, A. Perception Module · p. 3

A. Perception Module

  • A ROS 2 preprocessing node performs confidence-weighted RGB-D and 2D LiDAR depth fusion upstream, leaving the SLAM backend unmodified.The node publishes fused depth alongside the original color image for RTAB-Map ingestion.Source: 1) Preprocessing Pipeline · pp. 3–4

1) Preprocessing Pipeline:

  • The node pairs 30 Hz camera depth images with the most recent 10 Hz LiDAR scan using approximate synchronization with 50 ms tolerance.LiDAR points are transformed into the camera frame and projected using the D435i's factory-calibrated intrinsics.Source: 1) Preprocessing Pipeline · p. 3
  • Projected LiDAR depths are confidence-weighted with camera depths when both are valid; valid LiDAR depth alone fills invalid camera pixels.LiDAR returns are blended only in the camera–LiDAR overlap, and the resulting stream is published for RTAB-Map.Source: 1) Preprocessing Pipeline · pp. 3–4, Figure 3 · p. 4

2) FOV and Data Utilization:

The LiDAR scan is divided between depth fusion in the camera-overlap sector and a retained 360° scan stream for navigation. Sensor alignment uses a mechanically measured camera–LiDAR transform.

2) FOV and Data Utilization:

  • The LiDAR is cropped to an 80° sector for depth projection, while its remaining approximately 280° is retained for 360° obstacle detection.The fusion sector approximately matches the camera's horizontal field of view, with edge distortion accounted for.Source: 2) FOV and Data Utilization · p. 4
  • The LiDAR is mounted 15 mm above and 70 mm forward of the camera, with its rotation aligned to the camera's forward direction.The deployed mounting angle is CAD-calibrated and the translation offsets are mechanically measured.Source: 2) FOV and Data Utilization · p. 4
  • The sensor pipeline sends raw data through the fusion preprocessing node before RTAB-Map performs final map construction.The illustrated transform uses a nominal rotation of −90° and offsets of tx = −0.070 m, ty = 0 m, and tz = 0.015 m.Source: Figure 4 · p. 4, 2) FOV and Data Utilization · p. 4

3) Computational Benefit of Pre-Fusion:

Pre-fusion reduces the computational burden on the resource-constrained platform by sending one fused depth stream through RTAB-Map. The backend retains its existing graph-based optimization, while the reported loop-closure counts do not establish convergence or trajectory accuracy.

3) Computational Benefit of Pre-Fusion:

  • Pre-fusion replaces two odometry estimators and late fusion with one fused depth stream through a single visual-odometry pipeline, reducing computational load under an 8 GB shared-memory constraint.This architecture was chosen as a practical necessity for the Jetson Orin Nano.Source: 3) Computational Benefit of Pre-Fusion · p. 4

B. Graph-Based SLAM Backend

  • RTAB-Map optimizes pose-graph constraints, while the fusion node changes depth available for visual odometry and geometric verification of candidate loop closures.The fused stream produced a point cloud in which regions appearing as camera-only depth voids were filled by LiDAR-augmented depth.Source: B. Graph-Based SLAM Backend · p. 4, Figure 5 · p. 5
  • Indoor loop-closure counts describe effects in the recordings, not changes in optimizer convergence or absolute trajectory accuracy.The backend's odometry, loop-closure, and graph-optimization code is left unchanged.Source: B. Graph-Based SLAM Backend · pp. 4–5

C. Origami-Inspired Morphing Wheels

1) Mechanism Description:

Each origami-inspired wheel uses coordinated flaps driven by servos and gearing to shift between compact and extended profiles. The torque analysis supports the approximately 2 kg prototype, and milestone poses describe terrain and sensing configurations.

1) Mechanism Description:

  • Four servos drive each wheel’s four flaps through a spur-gear train engaging the flasher pattern’s ring gear, changing the wheel from a compact disk to an extended profile.Servo coordination controls the shape during the transition.Source: 1) Mechanism Description · p. 5

2) Kinematic and Static Analysis:

3) IMU-Compensated LiDAR Projection During Tilt:

Wheel-induced chassis tilt changes the 2D LiDAR scan plane, so the IMU rotation aligns returns to gravity for mapping. The platform integrates perception, navigation, and manipulation, although manipulation experiments are outside this paper’s scope.

3) IMU-Compensated LiDAR Projection During Tilt:

D. System Integration and Software Architecture

E. Implementation Details

The implementation combines an RGB-D camera, 2D LiDAR, and IMU with preprocessing upstream of RTAB-Map. Experiments characterize one prototype in an indoor textureless corridor and an outdoor sunlit area.

E. Implementation Details

  • Custom ROS 2 nodes perform depth fusion, coordinate alignment, and IMU-compensated tilt projection upstream of RTAB-Map, with a Madgwick filter estimating attitude.The system uses ROS 2 Humble on Ubuntu 22.04, with NumPy and OpenCV for sensor preprocessing.Source: E. Implementation Details · p. 6
  • The D435i provides depth reliably at approximately 0.3–3 m, with degradation beyond 3 m and depth voids on textureless surfaces and in direct sunlight.The camera supports up to 1280 × 720 at 90 fps.Source: E. Implementation Details · pp. 6–7, Figure 11 · p. 7
  • The RPLiDAR A2M12 provides depth from 0.2 m to its rated 12 m regardless of ambient lighting or surface texture; the fused stream defaults to an 8 m depth clamp.The clamp can be raised toward 12 m when longer range is required.Source: E. Implementation Details · p. 7

A. Experimental Setup

  • Experiments characterize a single prototype with a rigidly mounted D435i, RPLiDAR A2M12, and BNO055 IMU in the described locations.The modular sensor interface also allows the perception unit to be detached for handheld testing.Source: A. Experimental Setup · p. 7, Figure 10 · p. 7

B. Test Scenarios

  • The indoor scenario uses an approximately 10 m corridor with turns and a 3 m textureless wall; the outdoor scenario places the camera view across sunlit and shaded surfaces.The outdoor setting includes surfaces where reflected infrared returns can disagree with LiDAR range by more than the 0.5 m diagnostic threshold.Source: B. Test Scenarios · p. 7

C. Protocol

Each scenario used three manually piloted runs per configuration on comparable trajectories. The protocol defines frame-level void fraction and, outdoors, a camera–LiDAR disagreement statistic; these are descriptive depth-stream measures, not ground-truth benchmarks.

C. Protocol

  • Three independent runs per configuration followed comparable manually piloted trajectories, each lasting approximately 90–120 s.At 30 Hz, each run yielded about 2.7–3.6 × 10^3 depth frames.Source: C. Protocol · p. 7

1) Metric definition:

  • Per-frame void fraction counts pixels with zero or NaN depth over the image resolution, then averages across frames and runs.The resulting scenario/configuration values are descriptive statistics of the depth stream, not benchmarks against ground truth.Source: 1) Metric definition · pp. 7–8
  • Outdoors, the protocol additionally reports the share of LiDAR-projected pixels where camera and LiDAR depths differ by more than 0.5 m.This is a diagnostic disagreement rate, not a live fusion gate.Source: 1) Metric definition · p. 8

D. Origami Wheel Actuation

On one compact prototype, sensor fusion reduced invalid-depth fractions in indoor and sunlit outdoor tests, while wheel reconfiguration tilted the LiDAR through intermediate elevations. The section also reports mechanical cycling and indoor mapping observations, with no ground-truth pose-error estimate.

D. Origami Wheel Actuation

  • Adjacent wheel-pose transitions took approximately 1.5–2 s, with no flap or crease failure across 30 cycles on one prototype.The cycle test used the approximately 2 kg unloaded platform; Open extends wheel diameter for obstacle traversal, while Closed compacts the footprint.Source: D. Origami Wheel Actuation · p. 8

VI. RESULTS

B. Active Scanning During Wheel Reconfiguration

  • During wheel transitions, changing chassis pitch sweeps LiDAR scans through intermediate elevations; the most unfolded configuration reached approximately 14° pitch.The IMU-based projection accounts for attitude, but recordings provide no separate numerical estimate of volumetric coverage during transitions.Source: B. Active Scanning During Wheel… · p. 8

C. Mapping and Loop Closure

  • Indoor return-leg loop closure was accepted in all 3 fused-depth runs, versus 1 of 3 camera-only runs; the Jetson Orin Nano depth pipeline sustained 30 Hz with no observed dropped frames during recorded windows.Camera-only depth voids led RTAB-Map’s BoW geometric-verification step to reject candidate loop closures; outdoor loop closure was not recorded.Source: C. Mapping and Loop Closure · pp. 8–9

A. Sensor Fusion Results: Both Scenarios

  • Full-frame void fractions fell from 21% to 11% indoors and 48% to 18% outdoors; textureless-wall and sunlit-ROI fractions fell from 26% to 9% and 62% to 21%.These are depth-stream statistics averaged over three runs per configuration; the configured fused-depth limit was 8 m, versus an approximate 3 m camera reference indoors.Source: A. Sensor Fusion Results: Both Scenarios · p. 8, Table III · p. 9

VII. DISCUSSION

In the tested settings, RGB-D and LiDAR fusion reduced invalid depth, while wheel motion provided a changing LiDAR viewpoint. The discussion also notes limits in validation and unmeasured coverage during full wheel transitions.

VII. DISCUSSION

  • Indoor return-leg loop closure was accepted in three of three fused runs and one of three camera-only runs, but no ground-truth trajectories support absolute pose-error estimates.The authors characterize these depth-stream observations as motivation for further SLAM evaluation.Source: §VII · p. 9
  • During wheel reconfiguration, the LiDAR sweeps intermediate elevations; the IMU-reported pitch reached approximately 14° at the most unfolded position, but transition map-density gains were not measured.At held wheel positions, the scanner retains a fixed viewpoint.Source: §VII · p. 9

A. Limitations

  • Evaluation used one prototype, one indoor corridor, one outdoor location, and three manually piloted runs per sensor configuration, limiting generalization across tested settings.Trajectories were only approximately repeated, and wheel mechanisms, mounts, and extrinsic calibration may vary across builds; disagreement rates also depend on the 0.5 m diagnostic threshold.Source: A. Limitations · p. 9
  • The 2D LiDAR leaves much of the image outside its projected scan plane without range support.Source: A. Limitations · p. 9
  • Fusion and dense-grid generation consume substantial shared memory; payload approaches the prototype’s 2 kg design limit, and performance in rubble, smoke, water, low light, and rough terrain remains unmeasured.The reported memory use remains substantial even with 0.05 m voxel downsampling.Source: A. Limitations · p. 9

B. Future Work

  • Future work proposes using fused depth to pause or slow wheel transitions when more vertical samples could help resolve an obstacle, and comparing faster or higher-resolution LiDAR options.The authors also propose comparison with a servo-tilted 2D scanner or a native 3D LiDAR when payload permits.Source: B. Future Work · p. 9
  • Filtering transient returns before they enter the static map remains a separate challenge for future tests in dynamic indoor spaces and varied outdoor terrain.Source: B. Future Work · p. 9

VIII. CONCLUSION

  • Invalid full-frame depth fell from 21% to 11% indoors and from 48% to 18% outdoors.The conclusion reports these as the mean fractions in the indoor and outdoor tests, respectively.Source: §VIII · p. 9
Loading 2610.09828v1…