Hi HN, we're the founders of Salem Robotics (https://salemroboticsinc.com). We springiness existing mobile robots the task-specific intelligence to transportation retired surveys and physically interactive inspections successful hazardous business facilities.
Here's a video of it moving connected existent robot hardware pinch a fewer words from us: https://youtu.be/U_228h3NE7c
We came to Salem done robotics investigation astatine UT Austin and a mixed 15 years moving successful nuclear, including astir 10 years processing and deploying autonomous robots astatine Los Alamos National Laboratory. Over the past 5 years, we kept moving into the aforesaid gap: robot hardware had go very capable, but making a robot transportation retired a complete business process still required a astonishing magnitude of robotics activity and manual intervention.
The portion that willing america astir was manipulation. A atomic contamination survey, for example, tin require taking a "smear": wiping a defined area of a aboveground truthful it tin beryllium checked for removable radioactive contamination. In an oil, gas, aliases chemic facility, an LDAR (leak discovery and repair) inspection tin require moving a detector astir a peculiar valve, flange, aliases connection. Other inspections require positioning an instrumentality astatine a precise location and predisposition comparative to a tube aliases portion of equipment.
These are easy tasks to compress into verbs for illustration "wipe", "measure", aliases "inspect", but considerably harder to make a robot do reliably. A probe mightiness request to stay normal to a aboveground passim a path, enactment wrong a constrictive offset from a pipe, aliases trace a region while maintaining a peculiar end-effector orientation. The planner has to find a feasible mobility while respecting the task geometry, manipulator kinematics, associated limits, collisions, and the situation astir it.
We activity down to joint-level power for those interactions. One problem we've spent a batch of clip connected is generating constrained manipulation plans quickly capable that they tin beryllium based connected the geometry the robot really observes alternatively of requiring personification to cautiously writer a trajectory for each individual surface, valve, aliases flange.
The beingness world makes this annoying. A fewer centimeters of correction whitethorn not matter overmuch erstwhile navigating down a hallway, but it matters if a sensor is expected to stay normal to a curved surface. And successfully executing a trajectory doesn't needfully mean the inspection worked. The detector could beryllium misaligned, interaction could beryllium wrong, the geometry could disagree from the model, aliases the measurement itself could beryllium invalid. We attraction astir closing that loop astir the inspection result, not conscionable whether the limb reached the commanded pose.
Our attack is simply a operation of AI and classical robotics. A batch of robotics investigation and manufacture attraction correct now is going toward progressively end-to-end learned systems, peculiarly astir humanoids. Working successful safety-critical environments has made america admit really applicable classical approaches still are erstwhile you want definitive constraints, predictable behavior, and theoretical guarantees astir what a robot tin and cannot do.
We usage AI wherever semantic knowing and elasticity are useful, specified arsenic interpreting little system accusation aliases knowing what successful an unfamiliar segment is applicable to a procedure. Once the strategy knows what beingness relationship it needs to perform, we for illustration definitive geometry, planning, optimization, and power wherever possible. We're willing successful the matrimony betwixt the 2 alternatively than trying to make each portion of the robotics stack learned.
The different thought down Salem is that we don't deliberation each useful robot exertion should require building a caller robot. Companies for illustration Boston Dynamics are getting very bully astatine building progressively tin hardware platforms. We deliberation location is room for a domain-specific exertion furniture connected apical of that hardware. The aforesaid underlying robot mightiness execute atomic radiological surveys successful 1 installation and LDAR inspections successful another, but the procedures, sensors, manipulation constraints, occurrence conditions, and outputs are different.
That's besides why we're hardware agnostic. We don't expect 1 robot to beryllium the champion level forever, and accommodation already ain different hardware. We'd alternatively picture an inspection successful position of what needs to hap and past representation that onto the capabilities of the correct robot for the job.
One point that amazed america aft spending much clip pinch accommodation is really manual galore inspection workflows still are. In blase atomic and business sites, group still physically locomotion study routes, return measurements 1 astatine a time, visually inspect equipment, grounds results manually, and sometimes make judgments based connected things for illustration really a constituent sounds. Some of the basal workflows would beryllium recognizable to personification doing the occupation decades ago, moreover though the sensors, computation, and robots disposable coming are radically different.
We're starting pinch radiological inspection successful atomic because it's the manufacture we cognize best, and we're besides moving connected manipulation-heavy inspection problems successful oil, gas, and hazardous chemic facilities. The technicians and inspectors still specify the procedure, construe results, and make the consequential judgments. We're trying to automate much of the repetitive beingness execution that presently requires personification to participate the situation aliases manually run a robot.
We waste straight to business facilities, usually starting pinch a paid method validation and past moving to an ongoing deployment. Pricing varies substantially pinch the workflow: validations scope from tens of thousands to complete $100k, and larger deployments tin scope from the debased hundreds of thousands to astir $500k per robot.
One point we're particularly funny to perceive HN's thoughts connected is wherever the abstraction bound successful robotics should sit. What should travel from the robot manufacturer, what belongs successful an exertion layer, and what will inevitably stay circumstantial to the facility? We'd besides beryllium willing successful proceeding astir different industries wherever you've seen beingness inspection tasks that look trivial to a personification but are amazingly difficult to automate.
English (US) · 
Indonesian (ID) · 
