Modern robots don’t fail because they’re poorly trained — they fail because the real world is unpredictable. Unusual object placements, lighting shifts, human interference, and edge-case environments create failure modes no dataset can fully anticipate. When those failures happen in the physical world, they’re irreversible: collisions, damaged equipment, or unsafe human interaction.

Salus is Praxis Labs’ answer to this problem.

Salus is a real-time safety system designed to run alongside vision-language-action (VLA) models at deployment. Instead of trying to make models perfect at training time, Salus continuously monitors a robot’s behavior during execution, forecasts failures before they occur, and intervenes with safer alternatives in milliseconds. This enables prevention—not recovery—bringing the kind of predictive safety found in aviation and autonomous vehicles into embodied AI.

What makes Salus different is that it learns while deployed. Every near-miss, intervention, and correction feeds back into the system, allowing Salus to adapt to new environments, hardware differences, and real-world edge cases over time. As robots operate longer, Salus becomes more accurate, more efficient, and more context-aware—turning deployment into a safety learning loop rather than a static risk.

At Praxis Labs, we built Salus because we believe safe intelligence must be adaptive, not frozen. As robots move out of controlled labs and into hospitals, warehouses, homes, and public spaces, safety can’t be an afterthought or a training-time checkbox. It has to operate continuously, predictively, and transparently—exactly where failures actually happen.