The Robot That Can Move Is Useless If It Cannot See
A mining robotics company explains why perception is the hardest problem in industrial autonomy, and what happens when machines go blind in a dust cloud.
Key points
- Autonomous Solutions Inc. (ASI) has logged 4.5 million autonomous miles across mining, agriculture and construction sites.
- ASI's early automotive-grade sensors failed completely within a year: dust, vibration and temperature extremes destroyed them.
- Unplanned downtime at large mining operations can cost tens of thousands of dollars per hour, making perception failures extremely expensive.
- Modern industrial autonomy combines LiDAR for spatial mapping, cameras for visual context and radar for visibility through dust and fog.
A mining truck carrying 200 tons of ore doesn't get to stop and ask for directions. It needs to know, in under a second, that a light vehicle just crossed 80 metres ahead through a wall of red dust. Whether it can do that comes down entirely to how well it sees.
Mel Torrie, co-founder and CEO of Autonomous Solutions Inc. (ASI), a Utah-based company that builds autonomous systems for heavy industry, made that argument in a piece published by The Robot Report. His case is worth unpacking, because it addresses something most coverage of self-driving technology misses.
Why GPS alone is not enough
GPS tells a machine where it is on a map. Full stop. It cannot say what's directly in front of it.
For years, industrial autonomous vehicles leaned heavily on GPS for navigation. That works on a fixed route with no surprises. Real mine sites aren't fixed routes. Roads shift as extraction progresses. Obstacles appear without warning. In 45-degree heat, surrounded by dust, a satellite signal cannot keep a worker safe.
ASI learned this painfully. Early sensor kits borrowed from the automotive industry failed completely within twelve months. Dust and vibration destroyed them. The team had to design replacements from scratch, built for field conditions rather than car factories. As we reported on 20 July, Burro's co-founder made a similar point about agricultural robots: the lab isn't where hard autonomy problems live.
What does modern perception actually look like?
Three sensor types now work together on an ASI vehicle, each catching what the others miss.
| Sensor | What it does well | Weakness |
|---|---|---|
| LiDAR (laser mapping) | Precise 3-D distance measurement | Struggles in heavy dust or rain |
| CMOS camera | Reads visual context: person vs. Post | Needs light; fails in thick dust |
| Radar | Sees through dust, fog and darkness | Lower spatial resolution |
Running all three together means no single failure point blinds the vehicle. The system also processes sensor data on the vehicle itself, in real time, without sending anything to a distant server first. Waiting for a cloud response while a forklift crosses your path isn't an option.
The goal isn't a machine that stops when it detects an obstacle. A machine that simply stops halts a whole site. The goal is a machine that predicts what a nearby person is likely to do next and picks the safest response before the situation becomes a crisis.
Should ordinary workers be concerned about this technology?
For workers on industrial sites, better machine perception is directly protective. A system that can distinguish a person from a fence post and predict that the person is walking toward the vehicle's path is less likely to cause an accident than one that can't.
The bigger practical concern is integration. Many companies can't afford to scrap an existing fleet of expensive vehicles. ASI's answer is hardware-agnostic design, meaning the perception upgrade fits machines already on site rather than requiring full replacement.
The company says it has moved nearly 400 million tons of material across its autonomous deployments. Every trip, Torrie argues, taught them something new about what machines need to see.
Perception failures are the story here, not the motion-control advances that tend to dominate robotics coverage. The vendors selling autonomy on the strength of their actuators and gait algorithms are answering last decade's question. What matters now is whether the machine knows what it's looking at.
What to watch for: If your workplace is introducing autonomous vehicles, ask the vendor how the system handles sensor failure. A strong answer describes what the backup sensor layer does. A weak answer describes what the machine does when all sensors fail. Both questions matter.
