🐾Dynamic Obstacle Avoidance and Z-Axis Planning for Quadrupeds

Planning & Decision-Making

Carnegie Mellon University

Pittsburgh, PA

Aug 2024 - Dec 2025

🎯 Objective

Extend the Quad-SDK framework by integrating a dynamic obstacle avoidance system and Z-axis (height-aware) planning, enabling quadruped robots to navigate cluttered, dynamic environments more robustly.

💡Motivation

Quadrupeds are increasingly deployed in exploration, disaster response, and inspection tasks — where adaptability to dynamic, unstructured terrains is crucial.
Existing planners in Quad-SDK are limited to static 2.5D terrains and cannot react to moving obstacles or overhead barriers.

This project introduces real-time Z-axis adaptation and predictive obstacle avoidance, enabling quadrupeds to crouch or maneuver dynamically in cluttered spaces.

Challenges

RRT-Connect in Quad-SDK is unsuitable for real-time updates or dynamic obstacles.
Full 3D (x, y, z, velocity) planning rapidly increases computational cost.
2.5D terrain maps fail to capture 3D obstacle geometries.
Real-time motion prediction must balance reactivity and stability.

Gazebo Simulation

System Architecture

⚙️ Approach

1. System Architecture

Built on Quad-SDK, integrating modules for collision checking, Lazy PRM roadmap generation, and Weighted A* search.
Added Z-axis adaptation to allow crouching or elevation changes for obstacle traversal.

2. Collision Checking

Implemented bounding-box-based collision models and a sliding-window prediction mechanism for moving obstacles.
Predicts obstacle motion using velocity estimation and dynamically updates feasible paths in real time.

3. Planning Algorithm

Combined Lazy PRM with Weighted A* for efficient re-planning in dynamic conditions.
Lazy PRM retains invalid nodes with infinite cost to allow quick updates instead of full regeneration.
Weighted A* (ϵ = 1.7) accelerates search while maintaining sub-optimality bounds.
Integrated kinodynamic constraints (stance time, GRFs, accelerations) for physically valid paths.

Edge costs getting updated as obstacle is moving through the Lazy PRM roadmap

4. Z-Axis Planning

Introduced a lazy height adjustment mechanism to enable crouching or elevated walking.
If the nominal z collides, robot lowers to the minimum feasible z — otherwise continues at nominal height.
Achieves height-aware navigation without significantly increasing computation.

Plan generated by PRM-A*

Planning with z incorporated

📈Results

PRM + Weighted A* achieved ~20× faster planning than RRT-Connect (0.006 s vs. 0.01 s).
Lazy PRM enabled dynamic obstacle-aware updates without full graph regeneration.
Z-axis crouching allowed traversal under low-height obstacles while maintaining stability.
Enhanced planner maintained kinodynamically feasible, collision-free, and adaptive trajectories.

Dynamic obstacle avoidance.mp4