Certified guide dogs in Canada can cost tens of thousands of dollars to train, and applicants may wait several years before receiving one. Researchers at Mila, the Quebec AI institute, have built an alternative: Milo, a quadruped robotic guide dog that manufactures for under $2,000, navigates without pre-mapped routes, and runs entirely on onboard AI — no cloud computing, no internet connection required.
The full platform is open-source. Hardware designs, simulation software, and pre-trained AI models are all publicly available, which means any research team or engineer can download the plans, build the robot, and improve on the system without coordinating through a single commercial entity.
How Milo Navigates
Milo uses onboard AI to identify walkable paths, avoid pedestrians and obstacles, and guide users through unfamiliar environments. The navigation system was trained using reinforcement learning in a custom simulator that generated a wide variety of virtual environments and obstacle configurations — giving the system exposure to conditions that would be expensive to replicate in physical testing.
The robot interprets its surroundings using a bird’s-eye-view representation, which the developers say enables consistent navigation across different lighting conditions. Critically, Milo builds and maintains a local map of its surroundings as it travels, allowing it to remember obstacles that are no longer directly in sensor range. This persistent mapping reduces the risk of the robot forgetting a hazard it has passed and then routing back through it.
A telescopic handle connects the user to the robot physically. Sensors in the handle track the handler’s position and walking dynamics. A directional control pad lets users adjust walking speed and indicate preferred direction at intersections, preserving user agency over route decisions while leaving obstacle detection and pathfinding to the robot.
Why Open Source for a Guide Dog
The accessibility argument behind Milo is structural. The market for guide dog alternatives is large enough to matter — the World Health Organization estimates that at least 2.2 billion people globally have some form of vision impairment — but not large enough to attract the scale of commercial investment that consumer robotics categories receive.
By releasing Milo under an open model, the developers are attempting to bypass that constraint. Rather than waiting for a commercial company to build and price a product for a market that venture capitalists consider too small, any team with the capability can pick up the plans, improve the system, and deploy it in their region. The open-source release includes hardware schematics, simulation environments, and pre-trained models — everything needed to replicate and extend the platform.
The approach echoes the broader pattern in robotics where open-source platforms — like research access to Boston Dynamics’ ATLAS, or Nvidia’s Isaac platform — have driven faster iteration than proprietary systems. The difference is the target user. Most open-source robotics platforms serve researchers and engineers. Milo is designed for a user who is blind or visually impaired and needs a working navigation aid, not a research tool.
What the Testing Shows
The team evaluated Milo on indoor and outdoor obstacle courses, measuring navigation accuracy, travel speed, and obstacle clearance. The system has been tested in both controlled courses and real-world environments. The research paper, published alongside the open-source release, reports navigation performance across varied conditions.
What has not been tested at scale is long-term reliability, user acceptance, and performance in the full range of environments a real guide dog handles — crowded streets, public transport, weather, unpredictable pedestrian behaviour. Certified guide dogs undergo years of training and are matched to individual handlers. Milo, at this stage, is a research platform that demonstrates a viable technical approach to autonomous guidance at low cost, not a finished product.
The Cost Math
The sub-$2,000 manufacturing cost is the number that makes Milo worth paying attention to. A trained guide dog in Canada costs an estimated $40,000-$60,000 over its working life, with waiting lists of two to five years depending on the organisation. Robot guide dogs have existed in various forms for years, but they have typically been expensive, limited to indoor or pre-mapped environments, or dependent on cloud connectivity that limits their usefulness in areas with poor network coverage.
Milo’s combination of onboard autonomy, no-route-pre-mapping, and open-source availability at this price point is new. Whether it translates into real-world deployment depends on the same factors every assistive robotics project faces: durability, user trust, regulatory approval, and the gap between laboratory performance and the mess of everyday life.
❓ FAQ
Is Milo a replacement for a guide dog? Not yet. It is a research platform with open-source plans, designed to demonstrate that low-cost autonomous robotic guidance is technically feasible. The developers hope other teams will build on the platform to create deployable solutions, but Milo itself has not undergone the long-term user testing and regulatory clearance that a medical-grade assistive device requires.
How does it navigate without internet? Milo runs all AI processing onboard. It uses a camera-based bird’s-eye-view representation of its surroundings, builds a local map as it travels, and uses reinforcement learning models trained in simulation to identify walkable paths and avoid obstacles. No cloud connection is needed.
What does open-source mean here? The full platform — hardware designs, simulation software, and pre-trained AI models — is publicly available. Any researcher or developer can download the plans, build the robot, and modify or extend the system without licensing fees or coordination with the original team.
Where can I get the plans? The open-source release is available through Mila’s project page, alongside the research paper.
