MKMinkyung Jun
RESEARCH ยท Flagship master's project

SOAR-RL: Safe and Open-Space Aware
Reinforcement Learning Navigation

Period 2025.03 โ€” 2025.08 Team Solo (100%) Outcome First-author SCIE journal (Sensors) ยท ICCAS 2025
๐Ÿ“„ Paper (Sensors 2025) ๐ŸŽฌ Human detection demo ๐ŸŽฌ Obstacle avoidance demo

PROBLEMWhy this is hard

In corridors and narrow passages, a robot avoiding a person can easily corner itself against a wall โ€” or brush past the person threateningly. Classical planners treat people as mere moving obstacles and ignore where they are heading; existing crowd-navigation research assumes open spaces and degrades sharply in confined ones. This work starts from representing both "where people are going" and "where the robot can escape to" in the state.

METHOD 1Sensor-fusion human perception

The 3D LiDAR point cloud is preprocessed (ground removal, downsampling, ROI extraction), then matched with YOLOv11 person bounding boxes to cluster points per individual. Pedestrian velocity is estimated from frame-to-frame centroid motion, and a danger zone is generated dynamically for each person from position and velocity.

Three-stage pipeline: LiDAR preprocessing, RGB-LiDAR fusion, human clustering and danger-zone generation
Human detection pipeline โ€” (a) 3D LiDAR preprocessing โ†’ (b) RGB-LiDAR fusion โ†’ (c) per-person clustering, tracking, and danger-zone generation. Also published as two separate conference papers.

METHOD 2Human-Aware Occupancy Map (HAOM)

A high-resolution occupancy map (10 m ร— 10 m at 0.01 m/cell) around the robot fuses three layers โ€” Static (obstacles), Human (detected people), and DangerZone โ€” so the RL agent can distinguish "empty space" from "space a person is about to cross".

Diagram composing Static, Human, and DangerZone layers into the HAOM
HAOM composition โ€” Static + Human + DangerZone layers, also exposing open space.

METHOD 3Open-space-aware reward and training

In NVIDIA Isaac Sim / Isaac Lab, dynamic obstacles were placed in walled environments. The state combines HAOM obstacle information with traversable distance (open space); the reward exploits the open-space direction. Straight, L-shaped, and intersection courses with 1โ€“6 obstacles provided generalization across scenarios.

SOAR-RL reinforcement learning framework overview
DRL framework overview โ€” perception โ†’ HAOM โ†’ policy โ†’ control.
Intersection training scenario with robot and dynamic obstacles
Intersection scenario โ€” purple: dynamic obstacles, top right: goal.

RESULTSEvaluation

94%average success rate
(min 87%)
6%โ†“collision rate
(timeout 0%)
5.28smean travel time
fastest in comparison
40%+improvement over
classical ROS planners
MethodSR (%) โ†‘CR (%) โ†“TOR (%) โ†“ฮผPL (m) โ†“ฮผS (m/s)NT (s) โ†“
ORCA31.068.01.02.910.358.41
CADRL78.016.06.04.490.568.04
LSTM-RL89.08.03.07.050.878.10
DSRNN93.07.00.010.011.029.78
SOAR-RL (Ours)94.06.00.04.210.615.28

SR: success rate ยท CR: collision rate ยท TOR: timeout rate ยท ฮผPL: mean path length ยท NT: navigation time. Highest success rate with a 58% shorter path than DSRNN.

Rollouts over time in a six-obstacle environment ending in success
Rollouts with six dynamic obstacles โ€” red: robot trajectory, purple dots: pedestrians.

DEMOVideos

Official paper video โ€” training and evaluation across scenarios (3:52)

Human detection demo thumbnailโ–ถYouTube

Real-time human detection (3D LiDAR-RGB fusion)

Obstacle avoidance demo thumbnailโ–ถYouTube

Dynamic obstacle avoidance in narrow spaces

CITATIONBibTeX

@Article{s25175236,
  AUTHOR         = {Jun, Minkyung and Park, Piljae and Jung, Hoeryong},
  TITLE          = {SOAR-RL: Safe and Open-Space Aware Reinforcement Learning
                    for Mobile Robot Navigation in Narrow Spaces},
  JOURNAL        = {Sensors},
  VOLUME         = {25},
  NUMBER         = {17},
  ARTICLE-NUMBER = {5236},
  YEAR           = {2025},
  ISSN           = {1424-8220},
  DOI            = {10.3390/s25175236}
}