DebrisRL

Every year, people die trapped in burning buildings because structural collapse makes entry too dangerous for first responders. DEBRIS is a reinforcement learning agent trained to navigate collapsing, smoke-filled environments using only spatial sensors with no camera or visibility required.

Deployable on a ground robot or drone, it finds the path to safety and guides survivors out while firefighters coordinate from outside. No human enters the danger zone.

Built with PPO + MuJoCo physics simulation, trained across 4 progressive curriculum stages (static corridor → falling debris → scatter shrapnel → full chaos with ceiling crumble and fire zones). 4.3M training steps on Modal A100 GPU. The agent navigates 97% of a 28m collapsing corridor with only 1 avg collision, using a 71-dimensional sensor-only observation with no visual input.

The same sensory profile works in total darkness and smoke. This is exactly the conditions where cameras fail and people die.