Loading
0
Quadruped Locomotion   SpaceData Inc.
SpaceData Inc.   Internship Project   2025

Teaching robots
to walk through
the wreckage

A multi-terrain locomotion policy for quadruped robots, trained in Isaac Lab and tested in GPS-denied indoor environments built to simulate post-disaster scenarios.

Isaac Lab
RL Training
ROS 2 Nav2
Navigation
Isaac Sim
Simulation env
Quadruped
Platform

When a building collapses, the first 72 hours are when most survivors are found. Sending humans into unstable debris is dangerous. Wheeled robots get stuck. Drones can't carry sensors close to the ground. Legged robots can go where nothing else can.

The problem

Disaster environments
break every
assumption

After earthquakes, explosions, or structural failures, the operating conditions are unlike anything a standard robot is designed for. Floors are uneven. Corridors are blocked. There is no GPS. Lighting is unpredictable. And every second of delay matters.

Wheeled platforms lose traction on rubble and cannot climb over obstacles more than a few centimetres tall. Aerial robots struggle with endurance and cannot get close enough to debris piles to be useful. Legged robots are different. They can step over obstacles, maintain balance on uneven ground, navigate narrow passages, and carry sensor payloads deep into environments no human should enter.

The challenge is making them autonomous enough to do this reliably without constant operator input. That is what this project worked toward.

Operating conditions
Terrain
Debris and rubble
Irregular surfaces, step obstacles, narrow passages, unstable ground
Navigation
GPS denied
All localisation onboard using SLAM and sensor fusion
Visibility
Low and variable
Dust, smoke, damaged lighting, constrained sensor range
Time pressure
72 hour window
Most survivors found in the first three days after a structural collapse
Operator load
Minimal oversight
Emergency crews cannot dedicate attention to full teleoperation under field conditions
Platform
Quadruped robot
Four-legged platform, step climbing, dynamic balancing, sensor payload capacity
The multi-terrain policy trained during this internship is subject to an NDA with SpaceData Inc. and cannot be shown publicly. The video below demonstrates the flat terrain locomotion baseline that the work extended from.
Flat terrain locomotion baseline   Isaac Sim   ROS 2 Nav2
What I built

Multi-terrain
locomotion
from scratch

At SpaceData Inc. I was tasked with extending a flat terrain locomotion policy to handle the kinds of surfaces a robot would actually encounter in a disaster zone. That meant training a reinforcement learning policy in Isaac Lab that could maintain stable locomotion across rubble, slopes, steps, and irregular ground without pre-mapping the terrain.

  • Trained quadruped locomotion policy using RL in Isaac Lab
  • Multi-terrain curriculum across flat ground, rubble, slopes, and steps
  • Reward shaping for stability, smooth motion, and disturbance recovery
  • Policy outputs velocity commands compatible with ROS 2 Nav2
  • Navigation stack integrated with Isaac Sim for full autonomy pipeline
  • Disaster-inspired indoor environments built in Isaac Sim for validation
  • GPS-denied localisation using onboard SLAM
  • Deployment architecture designed to run on physical hardware
Status

Functional in sim.
Ongoing
towards hardware.

The multi-terrain policy reaches its navigation goals reliably across the simulation environments I built. The next step is validating it on physical hardware, tightening the navigation integration for tighter corridors, and improving recovery behaviors after unexpected falls.

Most of the detailed results from this internship are under NDA. What I can say is that this was the first time I had to build something with an actual operational constraint — not a research demo, but a system someone would actually deploy in the field. That changes how you think about every design decision.