The detail that jumped out at me most: after Nav2 kept failing to dock near the fridge, the robot wrote its own parking rule into places/rest-spot/SKILL.md, then parked there successfully using it. An agent patching its own navigation behavior is the kind of thing you hope this kind of stack can do, and here it actually happened.
The build itself: the agent is Claude Code running on an XLeRobot body, a black wheeled multi-tiered cart carrying a forward-facing camera plus planar LiDAR. Compute is an Orange Pi 5 Ultra with an optional LAN GPU box, and the SLAM layer, EvergreenSLAM, needs nothing more than a 2D lidar plus an RK3588 board. The lidar came in at roughly 30 USD, which keeps the sensing side cheap. Navigation runs on ROS 2 Humble with Nav2 using MPPI.
What stands out is how the agent talks to the map: a command-line tool called egs, where every location is a folder anchored to the lidar map, holding a notes.md file and often a photo. That structure pays off in the demo. Asked about scissors, the agent checked four known places and needed a hint on Day 1, but on a later run found them at the side table and sent a photo 44 seconds after reading the question. It also drove to the balcony on the strength of an inferred line in its own notes and photographed a spot it had never captured before. Booting on Day 2 at the parked spot with no relocalize command is a strong result too.
Not everything is closed off. The Day 1 scissors hunt took about ten minutes and four questions, the GPU box stopped responding mid-take, the agent sat on a message for around 90 seconds waiting for a listen call to end, and the demo included neither teleoperation nor arms. Six places saved in about six minutes of autonomous exploration on Day 1 is a solid start.
Project by Thomas. Watch it on their YouTube channel: