Skip to content
Early access openRev 0.1Robots in the lab 0001Local --:--:--
//Follow one weed

Laser weeder

A field machine that straddles the crop beds and kills weeds with lasers instead of herbicide. It creeps forward while cameras under a closed hood find each weed, and a laser burns the weed at its growing point.

Step 1 of 9: Light

Step 01 of 09, on the machine

Light

Inside the closed hood, LED bars light the bed the same way at noon and at dusk. A weed seedling a few centimetres tall comes into view between two rows of crop.

What it costs

If the exposure is 1 ms, the bed moves about 0.3 mm during it at 1 km/h. The lights have to be bright enough for exposures that short.

Estimate

Why it is hard, and what is at work

Sunlight through any gap overpowers the lighting and shifts plant colours. The hood has to seal against uneven ground without dragging through the crop.

Parts working at this step
  • Light-sealed hood

    A closed hood spans the beds. It shuts out the sun, so the cameras see every plant under the same light, and it keeps laser light and its reflections inside the machine.

  • Weed cameras and lighting

    Downward cameras beside each laser photograph the bed under steady LED light, and a model marks every weed.

Step 02 of 09, on the machine

See

The downward camera beside one laser takes a frame of the bed. The weed is in it, next to crop seedlings that look almost the same.

What it costs

A new frame every 25 ms, because the listed camera runs at 40 frames a second. At 1 km/h the bed moves about 7 mm between frames.

Derived from the Basler ace acA1920-40gc catalogue row

Why it is hard, and what is at work

A young weed and a young crop plant can look almost the same. Dust, wet leaves, soil type, and crop stage all change what the camera sees from one field to the next.

Parts working at this step
  • Weed cameras and lighting

    Downward cameras beside each laser photograph the bed under steady LED light, and a model marks every weed.

Step 03 of 09, on the machine

Decide

The computer on the deck runs the weed model on the frame, marks the weed, and finds its growing point, the few millimetres where a shot kills the plant.

What it costs

About 3 mm of travel for every 10 ms the decision takes at 1 km/h, roughly the width of the growing point.

Estimate

Why it is hard, and what is at work

The time from a camera frame to a laser firing has to be short and steady. A model that is fast on average and slow now and then misses the weeds that pass under the camera during the slow frames.

Parts working at this step
  • Compute and uplink

    A sealed computer runs weed detection and aims every laser, a CAN bus reaches the drives, and a cellular modem connects the machine to the people who run it.

Step 04 of 09, on the machine

Aim

The computer predicts where the weed will be when it reaches the laser row and turns the steering mirror in one laser module to that spot.

What it costs

About 2 s between seeing the weed and firing at it: the camera row sits about 0.55 m ahead of the laser row in this drawing, at 1 km/h.

Estimate

Why it is hard, and what is at work

The prediction rests on how far the machine has moved since the frame, measured from the wheels and from following the weed across frames. The wheels slip on soft soil, and slip too small for a driver to feel still moves every laser target.

Parts working at this step
  • Laser modules

    A row of laser modules, each aimed by a steering mirror across its own strip of bed, burns the growing point of each weed the cameras have found.

Step 05 of 09, on the machine

Fire

With the hood interlock showing the hood closed, the laser fires and the mirror puts the beam on the growing point. The weed is burned where it stands and the crop beside it is left alone.

What it costs

300 W of laser light while the beam is on, the listed laser's rated output, so every 100 ms on target delivers 30 J.

Derived from the IPG YLR-300-AC fiber laser catalogue row

Why it is hard, and what is at work

The growing point can be a few millimetres across, and the machine keeps moving while it fires, so the shot has to land where the target is at that moment. Every shot swings the load on the generator, and the supplies have to ride through it without resetting the computer.

Parts working at this step
  • Laser modules

    A row of laser modules, each aimed by a steering mirror across its own strip of bed, burns the growing point of each weed the cameras have found.

Step 06 of 09, across the fleet

Log

The machine records video of the bed in short segments, whether or not anyone is watching. Its own software writes down what the model decided about this weed and where the shot landed. Across 1,000 machines, records like these pile up faster than any team can look at them.

What it costs

About 92 MB a second of raw video from each camera: 1,920 by 1,200 pixels at one byte each before colour processing, 40 times a second.

Derived from the Basler ace acA1920-40gc catalogue row

Device to Cloud Flywheel solution: Cameras and vision devices

The fleet view shows 1,000 machines to show the problem at that size; Device to Cloud Flywheel has been run on fleets of single digits to low tens of devices.

Why it is hard, and what is at work

Recording everything fills the disk, and recording too little loses the one frame that explains a miss. What the machine keeps has to wait on its disk for as long as the cellular link is down.

Parts working at this step
  • Compute and uplink

    A sealed computer runs weed detection and aims every laser, a CAN bus reaches the drives, and a cellular modem connects the machine to the people who run it.

Where Device to Cloud Flywheel fits: Cameras and vision devices

Cameras in the field, fixed, on a machine, or a browser on a laptop. You want to see them live and pull frames back for training.

Demo

examples/browser-camera/demo.sh

Needs a running stack, gst-launch-1.0 for the synthetic camera.

The repository is private while the first release lands, so this demo cannot be run from this page. Write to hello@devicetocloud.ai and we will tell you when it opens.

From the solution page, punctuation adjusted.

Step 07 of 09, across the fleet

Return

When the link allows, the machine uploads its recorded segments. It also reports how it is doing. If its software runs on ROS 2, that report says which of its programs are running and which sensor feeds have stopped updating, without changing anything on the machine.

What it costs

At best, the listed modem sends 50 Mbit/s, about 6 MB a second, and a rural field usually gets much less. Even the best case is about 1 percent of what seven cameras record raw, so the machine has to choose what goes home.

Derived from the Sierra Wireless AirPrime MC7455 catalogue row

Device to Cloud Flywheel solution: A ROS 2 robot

The fleet view shows 1,000 machines to show the problem at that size; Device to Cloud Flywheel has been run on fleets of single digits to low tens of devices.

Why it is hard, and what is at work

Fields often have weak cellular coverage. With 1,000 machines, some of them are offline at any moment, so the server has to tell a machine that is quiet because it is fine from one that has stopped.

Parts working at this step
  • Compute and uplink

    A sealed computer runs weed detection and aims every laser, a CAN bus reaches the drives, and a cellular modem connects the machine to the people who run it.

Where Device to Cloud Flywheel fits: A ROS 2 robot

Robots already running ROS 2. You have a stack you are not going to replace, and you want fleet operations around it rather than instead of it.

Demo

embodiments/ros2-machine/demo.sh

Needs a running stack, uv. Drives the ROS 2 mock, not a live graph.

The repository is private while the first release lands, so this demo cannot be run from this page. Write to hello@devicetocloud.ai and we will tell you when it opens.

From the solution page.

Step 08 of 09, across the fleet

Learn

The fleet's records, this weed among them, are the raw material for the next model. People pick out the frames the model got wrong and label them, then send the set to a trainer. Device to Cloud Flywheel moves the files by checksum and hands each training job to the trainer, a machine you enroll like the weeders, when you start it. It doesn't choose the frames and it doesn't start training on its own.

What it costs

About 1 TB a day to sort if each of 1,000 machines sends home 1 GB.

Estimate

Device to Cloud Flywheel solution: Training a policy and shipping it to the fleet

A second product, Device to Cloud Engineer, is an engineer agent that writes the device code a machine like this runs. It is in development, and there is nothing to try yet.

The fleet view shows 1,000 machines to show the problem at that size; Device to Cloud Flywheel has been run on fleets of single digits to low tens of devices.

Why it is hard, and what is at work

The frames worth training on are rare: a new weed species, a different soil, a crop stage the model has not seen. At 1,000 machines the work is finding them in the pile and knowing which machine, field, and model version each one came from.

Where Device to Cloud Flywheel fits: Training a policy and shipping it to the fleet

Field data coming back from machines, and a model you want to improve and redeploy without a bespoke pipeline each time. It runs across the other solutions rather than describing one kind of device.

Demo

embodiments/train/demo.sh

Needs a running stack, uv.

The demo trains a small policy for a simulated excavator. It does not train a weed model. The loop is the same: files back, a training job you start, a checked file out to the machine.

The repository is private while the first release lands, so this demo cannot be run from this page. Write to hello@devicetocloud.ai and we will tell you when it opens.

From the solution page, wording adjusted.

Step 09 of 09, across the fleet

Ship

The new model goes out as a declaration of what each machine should run. Each machine downloads the file, checks it against its checksum, and loads it. In a fleet this size, a team gives it to a first group, compares that group's weed kills and misses with the rest, and then advances the rollout or pauses it and rolls back. On this machine, the computer judges the next weed with the new model.

What it costs

About 50 GB of downloads for one release, if the model file is about 50 MB and goes to 1,000 machines.

Estimate

Device to Cloud Flywheel solution: Training a policy and shipping it to the fleet

The fleet view shows 1,000 machines to show the problem at that size; Device to Cloud Flywheel has been run on fleets of single digits to low tens of devices.

Why it is hard, and what is at work

A model that is better on average can be worse in one field. The rollout has to pause on a regression, go back cleanly, and show which machines run which model, including the ones that were offline when it started. Deciding what counts as a regression is the team's call: it depends on weed kill rate and crop damage, which only the machine's own software measures.

Parts working at this step
  • Compute and uplink

    A sealed computer runs weed detection and aims every laser, a CAN bus reaches the drives, and a cellular modem connects the machine to the people who run it.

  • Weed cameras and lighting

    Downward cameras beside each laser photograph the bed under steady LED light, and a model marks every weed.

Where Device to Cloud Flywheel fits: Training a policy and shipping it to the fleet

Field data coming back from machines, and a model you want to improve and redeploy without a bespoke pipeline each time. It runs across the other solutions rather than describing one kind of device.

Demo

embodiments/train/demo.sh

Needs a running stack, uv.

The demo trains a small policy for a simulated excavator. It does not train a weed model. The loop is the same: files back, a training job you start, a checked file out to the machine.

The repository is private while the first release lands, so this demo cannot be run from this page. Write to hello@devicetocloud.ai and we will tell you when it opens.

From the solution page, wording adjusted.

And back

Back to 01 Light, with the new model

The next weed under the hood goes through the same steps. This time the computer judges it with the new model, and what the machine records about it feeds the model after that.

On the machine
Across the fleet
And back
//Every part in the cycle

Each subsystem has its own way to fail.

The catalogue examples are representative parts taken from vendor datasheets and product pages, each linked to its source. They show the kind of component each subsystem uses and the numbers that matter for it. They are not a bill of materials, and nobody has checked that they work together.

PartWhat it doesWhy it is hardCatalogue examples
Gantry frame and wheel drivesWhat it does. A welded steel gantry straddles five crop beds on four driven wheels, so the working tools ride above the crop while the tyres run in the furrows. Aluminium strut like the profile listed here mounts the lighter equipment on the deck.Why it is hard. The wheels have to hold the same furrows for hundreds of metres at a crawl, on soil that is soft in one place and baked hard in the next. Wheel slip too small for a driver to feel still moves every laser target, so the machine also follows each weed across camera frames.Catalogue examples
load capacity 1,996 kg · mass 95.7 kg · tread depth 24.6 mm
input voltage 24 to 80 V · output current max 350 A
profile width 45 mm · slot width 10 mm · mass per metre 1.6 kg
Light-sealed hoodWorking at 01 lightWhat it does. A closed hood spans the beds. It shuts out the sun, so the cameras see every plant under the same light, and it keeps laser light and its reflections inside the machine.Why it is hard. Sunlight through any gap overpowers the lighting and shifts plant colours, and a gap large enough for light to leave lets laser reflections out. The hood has to seal against uneven ground without dragging through the crop.Catalogue examples

No catalogue example listed.

Laser modulesWorking at 04 aim, 05 fireWhat it does. A row of laser modules, each aimed by a steering mirror across its own strip of bed, burns the growing point of each weed the cameras have found.Why it is hard. The growing point can be a few millimetres across, and the machine keeps moving while it fires, so each shot has to land where the target is at that moment rather than where the camera saw it.Catalogue examples
wavelength 1,070 nm · power 300 W · power tunability 100 %
power 500 W · wavelength 1,080 nm · modulation 50 kHz

The fibre lasers listed are examples of industrial laser sources, built for cutting and welding metal. A weeder chooses its laser wavelength for how well plant tissue absorbs it.

Weed cameras and lightingWorking at 01 light, 02 see, 09 shipWhat it does. Downward cameras beside each laser photograph the bed under steady LED light, and a model marks every weed.Why it is hard. A young weed and a young crop plant can look almost the same. Dust, wet leaves, soil type, and crop stage all change what the model sees from one field to the next.Catalogue examples
resolution 1,920 × 1,200 · frame rate 40 Hz · sensor size 1/1.2 in
Compute and uplinkWorking at 03 decide, 06 log, 07 return, 09 shipWhat it does. A sealed computer runs weed detection and aims every laser, a CAN bus reaches the drives, and a cellular modem connects the machine to the people who run it.Why it is hard. The time from a camera frame to a laser firing has to be short and steady. At 1 km/h, 10 ms of delay moves the target about 3 mm, roughly the width of the growing point. The field often has weak cellular coverage, so the machine has to keep working when the link drops.Catalogue examples
AI throughput 275 trillion operations/s · memory 64 GB · memory bandwidth 204.8 GB/s
downlink rate max 300 Mbit/s · uplink rate max 50 Mbit/s · transmit power 23 dBm
supply voltage 1.7 to 3.6 V · data rate max 8 Mbit/s · bus fault voltage 58 V

One embedded module is listed as an example. Seven cameras and seven lasers likely need several of them, or a larger computer.

Generator and power distributionWhat it does. An onboard diesel generator in a weatherproof enclosure feeds a distribution box that powers the lasers, the cooling, the computers, and the drives.Why it is hard. Each laser switches between idle and full power every time it fires, several times a second across the row. The generator and supplies have to handle those swings without a voltage dip that resets the computer.Catalogue examples
input voltage 90 to 264 V · output voltage 48 V · output current max 21 A

Seven 300 W lasers at about 30 percent efficiency draw about 7 kW on their own, before the cooling and the drives, so a real machine carries a generator of many kilowatts or draws power from a tractor. The catalogue has no generator row of that size. The power supply listed is an example of one stage inside the distribution box.

Chiller and radiatorWhat it does. A chiller pumps coolant through the water-cooled laser modules and the computer, and sheds the heat through a radiator.Why it is hard. The lasers and processors turn several kilowatts into heat in a hot, dusty field. The water-cooled nLIGHT laser listed under the laser modules is rated to run in surroundings from 10 to 40 °C. A radiator clogged with dust loses capacity gradually, until the lasers shut themselves down mid-row.Catalogue examples
temp −55 to 150 °C · accuracy max 0.1 °C · resolution 16 bit

The temperature sensor is an example of what watches the coolant. The catalogue has no chiller row. The IPG laser listed is air-cooled, so it would add its own heat to the air inside the machine.

Row guidanceWhat it does. Two satellite antennas on masts at the front corners give position and heading, an inertial sensor fills the gaps, and a forward camera watches the rows ahead so the machine steers along the beds.Why it is hard. Even with corrected satellite position, the rows sit where the planter put them, which a different tractor drilled on a different day. Steering has to follow the crop as it grew.Catalogue examples
accel range max 24 g · gyro range max 2,000 °/s · supply voltage 2.4 to 3.6 V

The inertial chip is an example sensor. Field steering uses a combined satellite and inertial receiver, and the catalogue has no receiver row.

Safety chainWhat it does. Emergency stop buttons at every corner, a hood interlock, and a warning beacon cut the lasers and the drives and warn anyone nearby.Why it is hard. A class 4 laser can blind a person with a reflection. The hood must be proven closed before any laser fires, and every stop has to cut power in hardware without waiting on software.Catalogue examples

The catalogue has no safety-rated part. A real hood interlock uses a certified two-channel switch wired to a safety relay and to the laser's own interlock input.

The drawing is illustrative, not a validated design. Costs marked estimate are estimates; the others are worked out from the catalogue part named under them.

//Where to start

The fleet half of this machine, as Device to Cloud Flywheel solutions.

  • Cameras and vision devices

    Cameras in the field, fixed, on a machine, or a browser on a laptop. You want to see them live and pull frames back for training.

    docs/solutions/cameras.md (opens when the code is published)

    In the cycle at 06 log. Quoted from the solution page, punctuation adjusted.

  • A ROS 2 robot

    Robots already running ROS 2. You have a stack you are not going to replace, and you want fleet operations around it rather than instead of it.

    docs/solutions/ros2-robot.md (opens when the code is published)

    In the cycle at 07 return.

  • Training a policy and shipping it to the fleet

    Field data coming back from machines, and a model you want to improve and redeploy without a bespoke pipeline each time. It runs across the other solutions rather than describing one kind of device.

    docs/solutions/train-and-deploy.md (opens when the code is published)

    In the cycle at 08 learn, 09 ship. Quoted from the solution page, wording adjusted.

None of the three solution pages has a section that separates measured results from claims, so nothing on this page is quoted as measured.

These pages live in the Device to Cloud Flywheel repository, which is private while the first release lands. Write to hello@devicetocloud.ai for access, and we will tell you when it opens.

//Where we help

Building one? Device to Cloud Flywheel is the software for the fleet half of this loop, from the first machine on.

Log, Return, Learn, and Ship: the steps after the machine works. The flywheel moves the data and the releases for every machine you have; what to train on and when to ship stay your team's call.