In the last post sprayer had milliseconds to decide and one nozzle to fire. Today the machine stops moving over the ground and starts removing it, and the clock relaxes while the geometry gets harder. An autonomous excavator digging a trench has to place the cutting edge of its bucket on a surface that exists only as a file: a design grade, a sloped plane a foot or two underground that no camera can see, because it is not there yet. Built Robotics quotes trench depths held to within a tenth of a foot, about three centimeters, with no operator in the seat and no survey stakes in the dirt. The thing that used to find that buried line, an experienced operator’s eye, has been replaced by arithmetic.

All week we have watched machines meet ground they did not design. A legged robot learned to feel its footing, a wheeled one learned when to stop rolling and climb, a drone paid an energy bill just to hover, a farm sprayer raced each plant through a narrow window. The excavator is the one that answers back. It does not adapt to the ground, it changes the ground to match a number. And to do that without a person watching the bucket, it has to know two things at once to within a few centimeters: where the design surface is, and where its own bucket teeth are. It has to keep knowing both while the soil pushes back. That last part is why digging to a spec turns out to be a harder problem than driving to a destination.

Think about what the machine is aiming at. A delivery robot drives to a point on a map, and a meter of slop is fine. An excavator digs to a design surface, a 3D model of the finished site that an engineer drew before anyone broke ground, and the allowed error is a few centimeters. The bucket has to land on that invisible plane on the first pass. Dig too deep and you have to haul in fill and compact it back up, which costs money and time. Dig too shallow and you come back and do it again. So the whole job comes down to one question, asked over and over: where is the bucket tooth right now, compared to the surface it is supposed to cut?

Answering it takes a chain of sensors that goes from coarse to fine. First, satellite positioning fixes where the machine is sitting, not to the few meters your phone manages but to a centimeter or two, using a technique called RTK that compares the satellite signal at the machine against a fixed station nearby. Second, a motion sensor on the body tracks how the machine is tilted and turned, so when it rocks on its tracks or sits on a slope, the system knows. Third, angle sensors on each joint of the arm, the boom, the stick, and the bucket, measure exactly how the arm is folded.

Now the geometry does the rest. If you know where the machine’s shoulder is and the angle of every joint below it, you can compute where the fingertip is. That is all the machine is doing: starting from a satellite-located point on its body and walking down the arm to figure out where the bucket teeth are, in the same coordinate system as the design. Subtract one from the other and you get a simple, useful number, how far the tooth is above or below grade. Computed many times a second, that number is the feedback loop a human used to close by eye.

Here is where two different products split apart. One is machine guidance: an in-cab screen, sold by companies like Trimble, Topcon, Leica, and CHCNAV, that shows a human operator exactly where the bucket sits against the model. With it, an average operator can hold grade like a veteran. The other is full autonomy, with nobody in the seat, which is what Built Robotics and Bedrock are building. Autonomy has to also do the part the human’s hands did, and that part fights back. A wheel rolling to a point is not resisted by the point. A bucket cutting earth meets soil that is hard in one spot and soft a meter away, and that resistance bends the arm and skips the tooth off line. The controller has to manage both the shape of the dig and the force of it, and the force is the part nobody can capture in a tidy formula.

There is a ruggedized, high-performance computer (often integrated directly into the touchscreen display inside the cab) that acts as the primary processor.

• What it does: This on-board computer takes the raw GPS/RTK coordinates, the chassis tilt data, and the arm joint angles, and runs the trigonometry calculations (sin and cos vectors) many times per second (usually between 20Hz to 100Hz, or 20 to 100 times a second).

• The Design File: The 3D blueprint or “design grade” file (usually a CAD surface or digital terrain model) is pre-loaded directly onto this in-cab computer via a USB drive or synced over a cellular connection beforehand. Because the design file lives locally on the machine, the computer can instantly subtract the bucket tooth’s real-time position from the design surface without any lag.

Before the data even reaches the cab computer, minor processing happens at the sensor level:

• RTK GPS Receivers: The GPS antennas on the machine handle their own complex calculations to compare incoming satellite signals with the corrections received from the local base station. It outputs clean, centimeter-accurate X, Y, Z coordinates to the cab computer.

• Smart Sensors: The angle and motion sensors (often IMUs, or Inertial Measurement Units) use internal filtering algorithms to smooth out high-frequency vibrations from the engine and tracks so they send clean angle data.

While the machine might be connected to the internet to report progress back to the office or receive design updates, it does not rely on a remote server or cloud software to calculate the real-time bucket position.

If it relied on the cloud, a temporary cellular drop or even a fraction of a second of network latency (ping) would mean the operator is digging blind. To catch a mistake before the bucket digs too deep, the feedback loop must be instantaneous, which requires local, on-board “edge” computing.

That is why this corner of robotics has turned to learning. A team at ETH Zurich trained a digging policy in simulation and watched it invent a sensible habit on its own: drag the bucket along the top in hard ground, push straight down in soft ground, switching by feel. Geometry tells the machine where to dig. The learned part is how to dig there when the ground argues back.

Bedrock Robotics raised $270 million in February, at a reported $1.75 billion valuation, to put autonomy kits on existing Caterpillar and Komatsu machines. Most of its founders came from the self-driving-car company Waymo, and its CEO says the first deployments with no operator in the cab are coming later this year. Low speeds and fenced job sites make construction a friendlier place to start than public roads.

Built Robotics, meanwhile, is already trenching on utility-scale solar farms, holding depth to a tenth of a foot and skipping the survey-and-grade-check crews that normally circle a dig. And for contrast, Dusty Robotics’ new layout robot prints building plans onto a finished concrete floor to about a sixteenth of an inch, far tighter than any excavator manages, simply because it works indoors on a flat, known surface. Across this whole field, accuracy is really a measure of how hostile the ground is.

The above visualization ranks four construction tasks by how tightly each holds position, from the indoor layout robot at a sixteenth of an inch, through machine guidance and the autonomous trench at a few centimeters, out to ordinary site driving where a meter is fine. The pattern is the quiet lesson of the week: the tight tolerances belong to the flat, controlled, indoor surfaces, and the excavator’s few centimeters are impressive exactly because they are won outdoors, in dirt, on a machine that rocks while it works.

So the excavator can hold a line it cannot see. The harder question is the one that opens next week, when the machine stands up and tries to look like us. A humanoid can do a flawless ninety-second demo on a stage and then fail an ordinary hour of real work. What separates the two? Next week we start with the hype, and with the single number that tells you whether a humanoid is actually doing a job or just performing one.

Subscribe for tomorrow’s read, we’re walking the robotics supply chain from atoms to algorithms, one weekday at a time.

Sources: