What to build

Build a working see → decide → pick system that finds fire hazards in a waste stream and physically removes them. Every project needs three parts:

1. See: a vision model that detects lithium-ion batteries, disposable vapes, power banks and battery-powered devices mixed in with ordinary waste. It needs to cope with clutter, dirt, partial occlusion and items that are crushed or half-hidden.
2. Decide: turn each detection into an action. Check that it's confident enough, pick up the most dangerous items first, and flag anything the model isn't sure about for a human to check.
3. Pick: a robot arm, gripper, pusher or air jet that actually gets the hazard off the belt, table or pile, in the real world or in simulation.

Any stack is fine. You can use real hardware, a low-cost arm (such as an SO-100/SO-101 with LeRobot) or a simulator (MuJoCo, Gazebo, Isaac Sim, PyBullet). A webcam, phone camera, depth camera or thermal camera all work. Some starting ideas:

• Belt sorter: a moving conveyor, real or simulated, where the robot has to pick hazards before they pass out of reach.
• Buried battery finder: fuse RGB with depth or thermal imaging to catch batteries hidden inside toys, cables and other e-waste.
• Garbage truck guard: a camera over the hopper that spots a hazard and stops the compactor before it crushes it.
• Drop-off assistant: a kiosk at a recycling center that identifies what someone drops in and sends batteries to a safe bin.
• Hard-case dataset: a new labelled dataset of difficult cases (crushed, dirty, half-hidden) plus a model trained on it and a robot that uses it.

Eligibility

• Solo builders and teams of up to 4.
• Anything started before the hackathon opened must be clearly marked, along with a description of the new work.
• Hardware is optional. Simulation-only entries are fully eligible.
• Safety first: never puncture, crush, heat or short-circuit a real lithium battery. For physical demos, use dead or dummy cells, vapes with the battery removed, or printed and 3D-printed stand-ins.

Submission requirements

• A public GitHub repo with a README that explains how to run it, what hardware or simulator it uses, and how the model was trained.
• Model results: your dataset sources, a held-out test set, and detection metrics (precision and recall, or mAP) for the hazard classes, including how many hazards it missed.
• A demo video (YouTube, Loom, Vimeo or X) of 3 minutes or less that shows the full loop running on a cluttered scene: the hazard detected, the decision made, and the hazard physically removed.
• A short description of what you built and why.
• No API keys in the repo. Use environment variables and include a .env.example file. Large model weights can be linked (for example on Hugging Face) instead of committed.
• One submission per person; you can update it until submissions close.

Judging

• Voting opens the moment submissions close and runs for 24 hours.
• Token holders vote, weighted by how much they hold at the snapshot (and still hold when they vote).
• The submission with the most vote weight wins. Ties go to the earlier submission.

Prize

• The winner receives the entire prize pool (sponsorship + 90% of creator fees earned so far).
• Every future fee claim is forwarded to the winner's payout address, forever.

Disqualification

Plagiarised work, non-functional demos or malicious code are removed from voting. So are:

• demos that are pre-scripted or teleoperated but presented as autonomous;
• detection-only entries with no pick, push or stop action;
• datasets used in breach of their licence;
• anything that damages, punctures or overheats a live lithium battery.

Resources

Vision

• Ultralytics YOLO docs
• Roboflow Universe (open datasets)
• TACO: Trash Annotations in Context dataset

Robotics and simulation

• LeRobot by Hugging Face
• MuJoCo
• Gazebo
• PyBullet