RV Lab | Mobile Robot

Perception · Mapping · Navigation

Mobile Robot

We build autonomous mobile robots that read people and terrain together—then turn uncertain sensor observations into safe, efficient motion in the real world.

RGB + 3D LiDARHuman-aware RLSemantic SLAM

88–94%SOAR-RL success rate

89%drivable-area IoU

3narrow-space scenarios

2025two journal papers

Research question

How can a robot find safe space before a crowded scene becomes dangerous?

Navigation in public space is not simply a shortest-path problem. Sidewalks narrow, pedestrians accelerate, and the boundary between drivable and forbidden space changes from moment to moment. Our work connects scene understanding, metric geometry, global mapping, and action selection in one deployable autonomy stack.

RGB cameras contribute semantics; 3D LiDAR contributes geometry; SLAM maintains spatial memory. Human-aware occupancy maps and open-space representations then give reinforcement-learning policies the information they need to move early and deliberately—not merely react at the last instant.

Core capabilities

One stack, from sensing to motion

Each capability is developed as a reusable module, but evaluated through the behavior of the complete robot.

01 / PERCEIVE

Human-aware perception

Fuse visual detections and point clouds to estimate pedestrian position, velocity, and situation-dependent safety regions.

02 / MAP

Traversable-space mapping

Project semantic drivable-area masks into metric space and accumulate them into a persistent SLAM-based map.

03 / ACT

Safe learning-based planning

Represent risk and open directions explicitly so the policy searches for viable corridors while progressing to the goal.

Research pipeline

Observe, remember, reason, move

A modular loop makes it possible to improve a sensor, map, or policy without rebuilding the entire system.

01 · SENSING

Multimodal input

RGB, LiDAR, odometry, and goal information are synchronized in the robot frame.

02 · REPRESENTATION

Semantic geometry

People, obstacles, road surfaces, and free directions become compact policy-ready features.

03 · LEARNING

Risk-aware policy

Simulation and reward design teach the robot to trade progress against future collision risk.

04 · DEPLOYMENT

Real-world validation

Classical baselines and changing layouts reveal failure modes beyond average performance.

Featured studies

See what happens inside the method

Paper figures now retain their natural proportions, so diagrams remain readable instead of being stretched into decorative banners.

SOAR-RL perception mapping and reinforcement-learning architecture

SOAR-RL · Sensors 2025

Planning with both danger and opportunity

SOAR-RL combines RGB–LiDAR pedestrian tracking with a human-aware occupancy map. A sector representation describes not only nearby obstacles but also the direction of open space, allowing the policy to commit to safer corridors earlier.

Evaluated in straight, L-shaped, and crossing narrow-space layouts
Compared with DWA, TEB, and trajectory-rollout planners
Reported 88–94% success across the three scenarios

Open paper and full-size figures →

Figure 1 from the linked open-access article (CC BY 4.0).

RGB LiDAR drivable-area recognition results in dynamic scenes

RGB–LiDAR Fusion · Sensors 2025

Drivable-area maps that survive sensor changes

The pipeline decouples RGB segmentation and LiDAR ground extraction before fusing them in a common frame. Because the components are modular, different sensor combinations can be used without retraining a monolithic fusion network.