The Problem: Disaster Search & Rescue
After earthquakes, explosions, or structural failures, first responders are often required to enter damaged buildings that are unstable, poorly lit, and unsafe for human access. These environments are cluttered, unpredictable, and typically GPS-denied — yet rapid situational awareness is critical.
Wheeled robots struggle with debris and elevation changes, while aerial systems face endurance and sensing limitations indoors. Legged robots offer a promising alternative, but enabling reliable autonomy in such environments remains a challenging systems problem.
- Collapsed or partially damaged indoor structures
- Narrow corridors, doorways, and stairwells
- Debris and uneven indoor surfaces
- GPS-denied localization
- Limited lighting and constrained visibility
What I Started Building
This project explores autonomous indoor navigation for legged robots in disaster-response scenarios, using simulation as a controlled environment to study system integration, failure modes, and design trade-offs.
Rather than treating locomotion and navigation as separate problems, the goal is to build a pipeline where learned legged motion can be reliably composed with classical navigation stacks to enable goal-directed autonomy.
Step 1: Learning Stable Legged Locomotion
The first step was to obtain stable and robust quadruped locomotion. Instead of hand-engineering gaits, a reinforcement learning policy was trained in Isaac Lab to learn walking behavior directly from interaction.
This approach enabled rapid iteration on reward design and produced smooth, disturbance-tolerant locomotion suitable for indoor environments with moderate irregularities.
- Reinforcement learning–based locomotion training in Isaac Lab
- Emphasis on stability, smooth motion, and recovery
- Designed as a foundation for higher-level autonomy
Step 2: Autonomous Navigation in Isaac Sim
With locomotion in place, the next step was integrating navigation on top of learned motion. Isaac Sim was used to construct indoor, disaster-inspired environments and connect the robot to the ROS 2 Nav2 navigation stack.
This integration surfaced practical challenges, including sensor synchronization, velocity command compatibility with legged motion, and navigation stability in narrow, cluttered spaces.
- Indoor navigation using ROS 2 Nav2
- Localization and planning in GPS-denied environments
- Bridging classical planners with learned locomotion
Current Status & Ongoing Work
This project is an active work in progress. While autonomous indoor navigation is functional in simulation, several challenges remain before deployment-ready autonomy can be claimed.
- Improving robustness under severe terrain variation
- More reliable perception in low-visibility environments
- Longer-horizon autonomy and recovery behaviors
- Future sim-to-real validation
The current system provides a strong foundation for continued exploration of legged autonomy in disaster response scenarios.