Problem
Indoor search-and-rescue is one of the harshest environments any robot operates in. Collapsed buildings after earthquakes, fire-damaged structures, post-blast or post-flood interiors — all share three properties that break most autonomous systems: no GPS (concrete, debris, partial roof collapse all block satellite signals), unstructured geometry (the building plan is no longer accurate; debris reshapes corridors hour by hour), and time pressure (golden hour for survivor location is roughly the first 24 hours, after which mortality rises sharply).
Current practice is human-led: rescue teams enter on foot or via dogs, often at significant personal risk, often slowed by the same conditions they’re trying to navigate. Larger drones exist for outdoor disaster response (mapping, thermal sweeps from altitude), but they don’t help once the search moves indoors.
Agitation
- Time matters more than perfection. A 70%-accurate map of a collapsed building delivered in 10 minutes is far more useful than a 99%-accurate one delivered in 4 hours. Rescue priorities — where are people likely to be alive? — change minute by minute as new information arrives.
- Risk to rescuers is real. Secondary collapses, gas leaks, electrical hazards in flooded basements — these kill or injure rescuers regularly. Sending a small drone in first to map and locate is a measurable safety improvement.
- Form factor matters. A drone that can fit through a 30 cm gap can enter spaces a human can’t. Most industrial drones are too large for collapsed-building interiors.
- GPS denial is the default. Every meaningful SAR drone application is GPS-denied. Solutions that depend on GPS+RTK don’t transfer.
Solution
A small indoor drone (30 cm motor-to-motor) with layered ToF sensing and an RL policy trained on cluttered environments. Specific properties that matter for SAR:
- Survives sensor degradation. The 3-rangefinder stack is layered specifically so that any single sensor’s failure modes (dust occluding the TF-Luna, smoke confusing the optical flow, dark surfaces fooling the VL53L0X) don’t take out the whole perception system. Critical when the air is full of debris.
- No-prior-map navigation. The RL policy is trained to explore unknown spaces — see the frontier-exploration alternatives doc for the planning side. SAR scenarios have no usable prior map; this matches.
- Thermal + visual payload. Same airframe, swappable payload. Thermal is what locates survivors in low-light or smoke-filled interiors; visual is what builds the map for the rescue team to follow.
- Cheap enough to lose. Sub-$1,500 BOM means deploying multiple units is feasible, and writing off a unit that doesn’t return from a collapsed structure is an acceptable cost. Industrial platforms in this price band don’t exist.
Operational picture
The realistic deployment is as a forward sensor, not a replacement for human rescuers. The drone enters first, builds a partial map, identifies thermal anomalies (potential survivors), and relays the data out. A human team then enters with a much better picture of where to focus. This is essentially the same loop military reconnaissance has used for decades, scaled down and indoor.
The hard remaining problems are comms relay (collapsed buildings block radio; we may need the drone to act as a mesh node or trail a fiber tether) and regulatory clearance for emergency deployment (current FAA/EASA frameworks don’t have a clean “emergency authorization” pathway for autonomous indoor drones, but precedent from disaster-response exemptions is forming).
What we’re not claiming
We’re not building a rescue robot that physically extracts survivors — that’s a different platform, much larger and very different engineering. We’re also not solving outdoor SAR, where established platforms already work well. The target is the indoor, post-disaster, time-pressured reconnaissance gap.
Where to go next
- Frontier exploration — how the drone handles unknown environments
- 3-rangefinder sensor stack — why redundancy matters in degraded environments
- Last-mile delivery — shared hardware, very different operational model