Introducing Mesos Robotics: Cleaning Up Our Cities with AI and Robots

The Problem

Walk through any city and the evidence is everywhere — cigarette butts jammed into gutters, plastic wrappers skittering across footpaths, cups and cans wedged under park benches. Urban litter isn't just unsightly. It blocks stormwater drains, fragments into microplastics, harms wildlife, and costs cities enormous sums to manage manually.

Existing solutions such as more bins, awareness campaigns, council clean-up crews can help at the margins, but they don't scale. The volume of litter entering public spaces every day vastly outpaces our ability to collect it by hand.

We think technology and AI can change that.

What We're Building

Mesos Robotics is building two things that work together:

A vision-language model (VLM) trained specifically for litter. Most off-the-shelf object detection models weren't designed to recognise the messy, varied, and often partially obscured items that constitute real-world litter. We're building and openly publishing a model purpose-trained on a large, richly labelled dataset of urban litter — across dozens of categories, lighting conditions, and surface types.

Autonomous robots that act on what the model sees. The VLM is only useful if something can do something about what it finds. We're developing a ground robot capable of navigating real city footpaths and parks, identifying litter in its path, and picking it up — operating safely and reliably alongside pedestrians.

These two systems are designed to complement each other, and we're building both from the ground up here in Sydney.

Open Research

We believe the environmental problem we're tackling is too important to keep behind closed doors. When our dataset and model weights reach a quality we're confident in, we'll publish them openly for anyone to use, fine-tune, or build on top of. If another team can use our work to solve a related problem - litter on beaches, in waterways, in rural areas. that's a win in our books.

How You Can Help

We're early stage and there's a lot of work ahead. The most valuable thing you can do right now is help us build the dataset that the model learns from. That means labeling images of litter — it doesn't require a technical background, just a bit of your time.

If you'd like to support the project financially, we accept donations that go directly toward hardware, compute, and research.

And if you're an engineer, roboticist, or machine learning researcher who wants to work on this, we'd love to hear from you.

We're just getting started. Thanks for being here.