Weeding Clawbot Videcoding Agents Kaggle Capstone July 2026

[00:00:03]It's one of me. The weeds don't take days off and paying a gardener felt silly when I could just invent one. This is the weeding claw bot.

[00:00:12]An AI agent that pulls weeds with a used 3D printer gantry and an Arduino.

[00:00:18]Built so the control layer refuses to target anything not classified as a weed.

[00:00:23]And unsure means don't grab. That refusal means "near is the hard part".

[00:00:28]And what it's what the project is about.

[00:00:32]Why an agent? Because well, robots are great factories because factories are built for robots, oddly enough.

[00:00:39]Gardener reason assigns itself weekly.

[00:00:42]Today's home weeders like Turtle just whack anything short and make you fence off your seedlings. The missing piece is knowing what you're looking at.

[00:00:51]An agent looks, decides, weed or flower, plans each grab, verifies its own work, retries failures.

[00:00:58]A human just has to get to that.

[00:01:02]The architecture is three layers of trust. A Gemini powered Gemini powered mission controller built with Google's agent development kit can only act through an MCP server exposing six tools. Detect, move, descend, grip, home, status.

[00:01:22]The raw G code, no serial access from the agent. Between them sits a safety safety layer.

[00:01:30]Envelope bounds, a flower guard that refuses any motion to the grab radius of a detected flower, an e-stop latch, and a motion budget.

[00:01:39]The gantry itself rejects physically invalid commands. The LLM The LLM is instructed, checked, and physically limited in that order.

[00:01:50]>> [snorts] >> Here's a real mission.

[00:01:51]Watch the tool calls. It's It surveys first, finds three weeds and two flowers.

[00:01:59]Then per weed, move, descend, close home to the bit.

[00:02:05]Grabbed every time.

[00:02:07]Then it scans to verify. It reports three weeds extracted, zero flowers or was damaged.

[00:02:14]The flowers never moved.

[00:02:16]And on a bed with only flowers, it detects issues, zero motion commands, and tells me the bed is clean. We have a dry check. Great test approves it.

[00:02:27]The hardware is real and cheap.

[00:02:30]SG90 gripper on an Arduino Uno R4. Here it is actuating.

[00:02:35]Riding a 3D printer gantry, that gantry is purely the demonstration prototype.

[00:02:40]The real machine is this frame on wheels. Starting at bed and rolling to the next. Swapping SIM for steel is what back end behind the same six tools.

[00:02:51]The safety guarantees and every test carried over unchanged.

[00:02:58]The build [snorts] itself was agent driven.

[00:03:01]Coding agents from the CLI to the implementation legwork. And here Andy Gravity is running a review pass, executing my test suite. 31 tests across three layers including trajectory tests that assert on the agent's actual tool call sequence, detect before move, descend, and before close.

[00:03:21]>> [snorts] >> Deployability is deliberately boring.

[00:03:25]Cloud had one API key, UV run. The lock file rebuilds the exact environment anywhere including the Raspberry Pi that will eventually live on the robot. The MCP server also speaks HTTP, so any client, curl, a dashboard, another agent can drive the gantry from the same six safe tools.

[00:03:45]Everything is open. Code, tests, the hardware, plan the future work.

[00:03:50]Next up, a teaching model where the robot reaches for each weed and waits for my go or no go and the autonomy bed by bed.

[00:03:59]Within FarmBot, agents you can trust around your flowers.