Heterogeneous Sensor-Based Multi-Robot Exploration for Radiation Source Localization in Unknown Environments (DARS 2026)

Smart Robots Design DEPARTMENT OF ROBOTICS TOHOKU UNIVERSITY Heterogeneous Sensor-Based Multi-Robot Exploration for Radiation Source Localization in Unknown Environments Akinori Watanabe1, Yusuke Tamura1,2, and Yasuhisa Hirata1 1) Tohoku University 2) Tokyo Denki University

Background Decommissioning after the Fukushima Daiichi nuclear accident •Radiation source locations and number are unknown after an accident •Debris and narrow passages restrict robot mobility •Teleoperation requires continuous supervision Current status of Fukushima Daiichi Nuclear Power Station [1] Autonomous source search is needed in unknown environments [1] T okyo Electric Power Company Holdin gs, Inc., “Status of Each Unit of the Fukus hima Daiichi Nu clear Power Station ,” https ://www.tepco.co.jp/en/hd/decommission/progress /about /i ndex-e.html 2026/11/4 DARS 2026 Smart Robots Design Lab. 1

• https://www.tepco.co.jp/en/hd/decommission/progress/about/index-e.html

Background Decommissioning after the Fukushima Daiichi nuclear accident •Radiation source locations and number are unknown after an accident •Debris and narrow passages restrict robot mobility •Teleoperation requires continuous supervision Inside the Unit 3 Reactor Building at Fukushima Daiichi NPS [2] Autonomous source search is needed in unknown, cluttered environments [2] T okyo Electric Power Company Holdin gs, Inc., “Internal Dose Investigation of the Unit 3 Reactor Building us ing a D rone”, https://p hoto.tepco.co.jp/en/date/2 018-e/2018 02-e/1802 28-01e.ht ml 2026/11/4 DARS 2026 Smart Robots Design Lab. 2

Prior Work: Using Dose-Rate Sensors Interpolation Dose-Rate Sensor Characteristics • Measures radiation intensity at each location • Small and lightweight for mobile robots • A single measurement does not indicate source direction • Mapping requires measurements at many locations Prior work using a dose-rate sensor (Abd Rahman et al. [3]) Dose-rate-only global search requires dense spatial sampling [3] Abd Rah man et al., "A coverage path plannin g approach for autonomou s radiation mapping with a mobile robot ", In ternation a l Journal of Advanced Robot ic Sys tems, 2022 2026/11/4 DARS 2026 Smart Robots Design Lab. 3

Prior Work: Directional Sensors Directional Sensor Characteristics • Measures the direction of incoming radiation • Narrows the search region with a few observations • Localization accuracy depends on measurement geometry • Shielding makes the device large and heavy Prior work using a directional sensor (Takahashi and Woo [4]) Source location is estimated from overlapping directional observations [4] Y. T akahashi and H. Woo, “Path Plannin g for Identification of Radi at ion Source Us ing Mobile Robot with D irection al Gamma -ray Detector”, IEEE/SICE Internatio nal S ympos ium on System Integrati on (SII), pp. 11 2–116 , 2 024. 2026/11/4 DARS 2026 Smart Robots Design Lab. 4

Objective  Most previous methods use only one sensor type  Sensor characteristics (✓ advantage, × limitation) Complementary Directional Sensor ✓ Directional data: efficient global search Dose-Rate Sensor ✓ Small size: high mobility Scalar data: low search efficiency Large size: low mobility Autonomous radiation source exploration in unknown environments with directional and dose-rate sensors 2026/11/4 Smart Robots Design Lab. 5

System Overview  Heterogeneous Robot Team  Small Robots  Large Robot • 2026/11/4 DARS 2026 Generates global source candidates from directional measurements • • • Explore candidates with dose-rate sensors Refine locations and reject ghosts Shared Map and Candidate List Smart Robots Design Lab. 6

Directional Candidate Generation and Ghosts Ghost Superimpose directional distributions Use local maxima as candidates Crossing directions from multiple sources create false maxima (ghosts) 2026/11/4 DARS 2026 Smart Robots Design Lab. 7

Candidate Allocation and Navigation Market-Based Task Allocation Bid𝑟,𝑗 = 𝛼𝑤𝑗 − 𝛽𝑑𝑟,𝑗 − 𝛾𝑐𝑗,𝐾 + 𝜖𝑗 normalized importance of candidate point 𝑗 normalized Euclidean distance from robot 𝑟 to candidate point 𝑗 small perturbation used only for tie-breaking normalized congestion term •Reallocate using current robot positions and the shared map •Remove unreachable and processed candidates •Replan with D* Lite when a new obstacle is detected Replan with D* Lite 2026/11/4 DARS 2026 Smart Robots Design Lab. 8

Local Exploration and Position Refinement Local Candidate Verification 1 Measure a 3 × 3 local region 2 Track the raw-data GPR peak 3 Estimate the final position with log-GPR Ghost Rejection Step limit reached or Estimated peak intensity below the threshold 2026/11/4 DARS 2026 Smart Robots Design Lab. 9

Experimental Setup and Evaluation Scope Main Conditions •50 × 50 grid, 1 m per cell •Two small robots with dose-rate sensors •The robots can communicate with each other Large-Robot Assumptions  Four directional measurements are prescribed  Large-robot motion, accessibility, and acquisition cost are excluded Environment setup Coverage and travel distance evaluate only the small-robot verification stage 2026/11/4 DARS 2026 Smart Robots Design Lab. 10

Baseline  Using Dose-Rate Sensors Only  Coverage-based exploration (Hazon et al. [5])  Visit every free cell Coverage 100% Total Travel Distance 2311 m Without candidate guidance, the robots must cover the whole environment Robot Trajectories for the baseline [5] Hazon et al., “Towards robus t o n-line multi-ro bot c overage”, Proceedi ngs of th e IEEE In ternational Conference on Robotics and Automation 2026/11/4 DARS 2026 Smart Robots Design Lab. 11

Experimental Cases Case Case 1 Case 2 Case 3 Case 4 Number of sources 2 3 4 4  Cases 1-3 The number of sources increases from two to four  Cases 3 and 4 Both contain four sources, but use different source arrangements Case 3 is the representative example shown in the following slides 2026/11/4 DARS 2026 Smart Robots Design Lab. 12

Proposed Method: Case 3  Initial Setup  Four true sources  Four directional measurement positions  Two small robots at the baseline start positions  Source locations are unknown to the robots Environment Configuration for Case 3 2026/11/4 DARS 2026 Smart Robots Design Lab. 13

Proposed Method: Case 3  Directional Candidates  White circles: source candidates  Black crosses: true sources for evaluation  Candidates appear near the true sources  Ghost candidates also appear 2026/11/4 DARS 2026 Smart Robots Design Lab. 14

Proposed Method: Case 3  Multi-Robot Exploration  Gray: unknown area  White: explored free space  Yellow: candidate points  Two robots verify different candidates in parallel 2026/11/4 DARS 2026 Smart Robots Design Lab. 15

Proposed Method: Case 3  Final Results •All four true sources localized •Ghost candidates rejected •Two small robots share candidate tasks •Mean localization error: 0.083 m 2026/11/4 DARS 2026 Coverage 39.8% Total Travel Distance 634 m Smart Robots Design Lab. 16

Comparison with total travel distance Baseline 2311 four cases  Travel distance reduced by 63.1–96.3%  Only 8.0–52.2% coverage required Total travel distance [m]  All true sources localized in all Case 4 852 Case 3 634 Case 2 675 Case 1 85 0 2026/11/4 DARS 2026 Smart Robots Design Lab. 500 1000 1500 2000 2500 17

Summary Results Evaluation Scope •Combined directional and dose-rate sensing for autonomous source exploration •Communication limits and failures are not evaluated •Detected all sources in four cases and reduced mean error to 0.083-0.141 m •Large-robot motion, accessibility, and acquisition cost excluded •Other environments and physical robots remain future work Global candidate generation and local refinement enable accurate localization with limited exploration 2026/11/4 DARS 2026 Smart Robots Design Lab. 18