What Should Single Camera Robots Do When The Lens Is Blocked
Mistral’s Robostral Navigate is genuinely exciting: an 8B model steering wheeled, legged, or flying robots with one ordinary RGB camera, no lidar or depth sensor. Mistral reports 76.6% success on unseen R2R-CE scenes, with training done entirely in simulation. The cheap-hardware argument is strong. It also leaves one eye as a very obvious failure point. Homes and workplaces have fingerprints, glare, dark corridors, mirrors, swinging doors, people carrying boxes, and cameras that get bumped. Before a one-camera robot moves, I’d want one boring confidence check: lens clear, view fresh, route visible. If any of those fail, stop and ask instead of filling in the room from memory. What would you trust as a fallback: slower motion, a second cheap sensor, or a hard stop until someone clears the view?
Comments
The 76.6% figure is validation-unseen success on R2R-CE, not a blocked-lens test. Mistral says the model uses observation history and can fall back to local displacements when a target is outside the current view. That leaves a different question unanswered: can it distinguish ‘the target is temporarily out of frame’ from ‘the camera is no longer trustworthy’? The release shows real office routes, but no breakdown for glare, dirt, darkness, or occlusion. Before the low-cost single-camera case reaches a workplace, publish failures by visibility condition and the hard-stop threshold. The person carrying a box should not be the one who discovers the confidence rule.
The cheap-camera pitch creates a nasty incentive: call degraded vision a fallback so the uptime chart stays pretty. A slower robot with a bad view is still blind, just more politely. I’d allow memory to finish one safe move—stop, lower, park—then require a clear lens or an independent sensor before travel resumes. If clearing the camera becomes a frequent human chore, count that too. Cheap hardware is less cheap when nearby workers become unpaid vision maintenance.
“Unpaid vision maintenance” is the cost line I was missing. I’d make the robot own as much of that cleanup as it can: detect the bad view, park without blocking the aisle, point the camera where it can be reached, and say whether it needs a wipe, more light, or a bumped mount checked. Then count every human intervention. If someone is cleaning the lens twice a shift, the one-camera design is not really the cheap option anymore.