A new navigation framework called MVP-Nav shows that robots can move safely through physical spaces without ever needing a depth sensor.
Researchers introduced MVP-Nav to tackle zero-shot object goal navigation — the task of finding an object in an unfamiliar environment with no prior training on that specific space. The system takes ordinary RGB camera frames and uses 3D foundation models to infer physical geometry, projecting 2D images into 3D bounding boxes that map out what is actually solid. It then feeds that reconstructed geometry into what the paper calls a Multi-layer Value Map, a shared cost space that weighs semantic reasoning — "that looks like a chair" — against physical constraints — "that chair is behind a wall I cannot pass through." Benchmarks show MVP-Nav outperforms existing depth-free methods on standard zero-shot navigation tests.
The gap this closes matters because depth sensors are expensive, power-hungry, and fragile — the kind of hardware that makes consumer robots cost-prohibitive. If a vision-only system can match depth-sensor navigation reliably enough, it lowers the floor for deploying embodied AI in homes and warehouses where LIDAR rigs are not practical. The harder problem MVP-Nav names — semantic-physical misalignment, where a model knows what something is but not where it physically sits — has quietly undermined a lot of promising navigation research.
It is worth noting this is a preprint, not a peer-reviewed product, and lab benchmark performance has a long history of not surviving contact with cluttered living rooms.