ā† Back to Projects

Dynamic Obstacle Detection and Avoidance

Dynamic-obstacle avoidance on a Husky A200 using multimodal perception and reactive planning

Akash Kumbar, Astik Srivastava, and Nabanita Dash

System
RGB-D/LiDAR obstacle tracking, Fast-LIO localization, RRT*/MinCo global planning, and MPC-based reactive avoidance.
Stack
ROS Noetic, RViz, Fast-LIO, onboard-detector, RDA planner, Ubuntu 20.04
Hardware
Husky A200, Intel RealSense D455, Livox Mid-70, Intel i5 + RTX 2060
Course Report (PDF) ā†— Real-world Robot Demo ā†—

We developed a navigation stack for the Clearpath Husky A200 to detect and track moving obstacles, localize against a LiDAR map, and reactively plan collision-free motion. The demonstrated system uses a prebuilt LiDAR map for localization; moving obstacles are detected and tracked online. Our team integrated the perception, localization, and planning modules through ROS topics and services, using RViz to inspect sensor overlays, robot poses, and trajectories.

Husky deployment, as reported in the course report, Sections 5 and 6.
MeasureReported resultEvaluation context
Collision avoidance95% of dynamic-obstacle encountersMore than 20 test trials in indoor and outdoor settings, including moving humans. Success refers to avoiding a collision during an encounter.
Full planning-loop rateApproximately 0.8 HzAbout 1.25 seconds per planning update. The report does not give a separate low-level control frequency.
Onboard computeIntel i5 + RTX 2060ROS Noetic on Ubuntu 20.04, with RGB-D and LiDAR perception on the physical Husky A200.

The report gives the aggregate success percentage and a trial-count lower bound, but no exact encounter count or per-trial breakdown. The deployment uses a prebuilt map; online map updates are listed as future work. Read the evaluation ā†—

System overview

System architecture for perception, localization, planning, and execution
Perception, localization, global planning, reactive planning, and robot execution.
LiDAR and visual perception pipeline
Multi-modal perception and dynamic-obstacle tracking pipeline. Image taken from LV-DOT.

Dynamic actor detection and tracking

Detecting and tracking a moving actor using the onboard perception stack.

LiDAR localization

The robot localizing with LiDAR against a provided map.

Reactive Planning and Real-world Navigation

RViz visualization

RViz view of localization, obstacle perception, and robot motion. Open in Google Drive ā†—

Real-world Husky demo

Real-world robot navigation and reactive obstacle avoidance. Open in Google Drive ā†—

Husky platform

Clearpath Husky A200 with RGB-D camera and LiDAR
Hardware setup on the Clearpath Husky A200.
Husky robot model in simulation
Husky platform in simulation.