Salem Robotics is trying to solve the "last mile" problem for
If you've ever worked with mobile robots, you know that navigation is relatively solved. A robot can drive down a hallway without hitting a wall. However, the moment you ask that robot to perform a "smear test" in a nuclear facility—where it has to wipe a specific surface to check for contamination—or to maneuver a leak detector around a complex valve in a chemical plant, everything breaks.
The manipulation gap
The founders of Salem come from a heavy-duty background, including research at UT Austin and a decade of deploying autonomous systems at Los Alamos National Laboratory. They noticed a recurring pattern: hardware has become incredibly capable, yet performing a complete industrial procedure still requires an absurd amount of manual intervention and custom robotics engineering.
The difficulty lies in the "verbs" of industrial work. It's easy to say "wipe this surface" or "inspect this flange," but translating that into machine motion is incredibly complex. A probe needs to stay perpendicular to a curved surface while moving, or maintain a precise offset from a pipe while navigating around obstacles.
Why end-to-end AI isn't the answer here
There is a huge trend right now in the AI community toward end-to-end learned systems, especially with the rise of humanoid robots. While that's fascinating for general-purpose tasks, Salem is taking a much more pragmatic, hybrid approach. When you are working in safety-critical environments like nuclear or oil and gas, "it usually works" isn't a good enough metric. You need explicit constraints and predictable behavior.
Their AI workflow looks like this:
- AI for Semantic Understanding: They use AI to interpret unstructured data and figure out what objects in an unfamiliar scene are actually relevant to the task (e.g., "That's the valve I need to inspect").
- Classical Robotics for Execution: Once the target is identified, they switch to explicit geometry, planning, optimization, and joint-level control.
This allows them to generate constrained manipulation plans based on what the robot actually sees in real-time, rather than requiring a human to pre-program a specific trajectory for every single valve or pipe in a facility.
Closing the loop on inspection
One of the most interesting technical points they made is that reaching a commanded pose doesn't mean the job was done correctly. In a real-world deployment, the sensor might be misaligned, the contact pressure might be wrong, or the physical geometry might differ from the digital model. Salem's goal is to close the loop around the result of the inspection, not just the movement of the arm.
Instead of forcing companies to buy entirely new, specialized hardware, they are building the intelligence layer that can be deployed on existing mobile robots. It's a smart way to tackle the deployment hurdle in industries that are notoriously slow to adopt new tech.