An Open-Source Robot Guide Dog Undercuts the $50,000 Real Thing

A new arXiv preprint undercuts the roughly $50,000 cost of a trained guide dog with a fully autonomous, open-source robotic platform built on a roughly $2,000 hardware stack. The release publishes the complete software alongside the hardware, the part that turns a single research project into a community-rebuildable artifact.

The platform is called Milo, and it lands in a category, assistive robotics for blind and low-vision (BLV) users, where the only mature option, a trained guide dog, runs about $50,000, takes years to match with, and works for roughly eight years before retiring. The waiting lists are long enough that many applicants never reach the top. The replacement rate is short enough that mobility comes with a renewal clock.

The Milo paper is the first to release both the hardware build and the full software stack in one place. A blind or low-vision user, a researcher, a community robotics lab, or an accessibility-focused maker space can now rebuild the same platform, modify the perception stack, swap the navigation policy, or add sensors, instead of licensing a vendor's black box. The robot itself is a modified Unitree Go2 Air/Pro quadruped (a consumer four-legged base) with a LiDAR sensor, a front-facing RGB camera, a telescopic guide handle, and a 2-DoF joint with magnetic encoders that track the handler's position. An Nvidia Jetson Orin Nano runs everything onboard, and no a priori 3D map is required.

The mechanism inside the paper is the part that determines whether the platform matures. Milo builds a bird's-eye-view (BEV) representation of the floor, obstacles, and pedestrians from RGB, LiDAR, odometry, and handle-encoder input, and feeds it to a reinforcement-learned obstacle-avoidance policy trained in a custom simulator. A safety filter wraps the policy output before commands hit the motors. The authors report that the learned policy produces smoother navigation and fewer handler collisions than a costmap-based baseline (a traditional geometry-only planner) in their indoor and outdoor obstacle-course tests, and that Milo is the only system in their comparison table to satisfy seven criteria at once: indoor and outdoor operation, operation in an unknown environment, fully onboard compute, handler commands, dynamic handler modeling, and open-source release.

The honest limit is the evaluation scope. The tests are "real indoor and outdoor obstacle courses," not day-to-day guide-dog work: no grocery runs, no crowded transit platforms, no weather. The RL policy is trained in simulation, so generalization to messy real-world conditions is the next open question, not a settled answer. The platform is a research contribution and a candidate for accessibility, not a deployed consumer product, and no independent BLV-community validation, user study, or third-party review is in the preprint or project materials.

The open release enables a different iteration loop. Trained guide dogs are a closed product: the breeding, training, and matching pipeline is owned by guide-dog organizations, the cost is fixed, and the pace of change is set by the organizations, not the users. A research platform that hands the perception, navigation, and hardware design back to the BLV and robotics community inverts that dynamic. The next iteration, a better handle sensor, a cheaper LiDAR, a community-trained policy that handles a specific intersection, comes from whoever builds it, not whoever sells it.

The hardware bill of materials and software stack are both published on the project site. The first community fork, not the original authors' results, will decide whether Milo becomes a real mobility tool.