Neufangled Sells Stringman, a $795 Open-Source Ceiling Robot That Tidies Rooms

TL;DR

Stringman is a room-scale robot with no wheels, no legs, and no battery. Four motorized anchors mount near your ceiling corners, spool lines down to a two-finger gripper, and that hanging gripper can carry up to 2 kg to any point in the room. Neufangled Robotics, an open-source hardware startup founded in 2025 in North Carolina, sells the assembled version for $795, publishes the firmware, PCBs and CAD, and ships it compatible with LeRobot so you can train new behaviors from about 50 teleoperated episodes. It reached the Hacker News front page on November 3 with 619 points, and the whole AI stack runs on your own machine.

The bet: work from the ceiling

Founder Nathaniel Nifong's insight is about cost. "In robotics, price is basically actuator count," the product page argues, so Stringman reaches an entire room with four motors instead of the seven a desk arm needs. The motors are Damiao H6215 brushless direct-drive hub motors with integrated FOC controllers, two per anchor, with two passive eyelets completing four pull points. The site's slogan for the architecture is "Tension is all you need."

Cables change the payload math too. Near the center of a 7 x 7 x 3 m room the lines share the load like four people carrying a couch, so their torque adds together instead of one joint fighting gravity alone. Average payload over the work volume is 1.9 kg, against 0.32 kg for a 7-actuator arm built from the same motors, by the maker's calculation. Work area is 147 cubic meters, about 500 times the volume per dollar of a $2,500 research arm, at comparable payload.

The geometry is old news to anyone who has watched a stadium broadcast: Skycam has flown cameras on four cables over football fields since the 1980s. What is new is selling that kinematics as a domestic product, with an OSHWA-certified open hardware license and a price under $800.

What $795 gets you

The assembled kit covers two anchors, the Arpeggio gripper with a wrist and two fingers, and a wireless hardware e-stop. Sensing is three cameras: two 1920x1080 anchor cams at 30 fps for global localization and a 684x384 gripper cam at 60 fps aimed at the fingers. Onboard compute is deliberately minimal, a Raspberry Pi Zero 2 W in the gripper and a Pi 3A+ at the anchors, streaming video to a host PC over Wi-Fi. You supply that host: any Linux or Mac box with 6+ cores, and a GPU or NPU if you want AI control, with full support for Apple Silicon. Rooms from 3 to 10 m diagonal with 2.3 to 4 m ceilings work, and the maker reports about 1,000 hours of testing in real homes so far.

Safety lives in the mechanics and the electronics. Any line that exceeds 28 newtons of tension, about the weight of a backpack, makes its motor release with damped motion, so a person or a ceiling fan wins every contest. Nifong put the same point more directly in the Hacker News thread: "Can it kill you? no probably not."

How the models drive it

Stringman's default brain is a hybrid: a custom vision model picks targets, LeRobot-trained policies do the grasping, and printed April tags on bins mark drop-off points. The networks have learned that the word "toybox" means the tag, so you can point the robot at containers that look nothing like its training data. Out of the box, the one fully autonomous behavior is picking up laundry and dropping it in the hamper, which is the hello world of home robotics with unusually good reach.

Adding a behavior of your own follows the standard imitation-learning loop, minus the lab. You teleoperate with a gamepad from the web UI, record episodes from that same UI, train a policy in Google Colab with LeRobot, and run it back on the robot. The docs say as few as 50 episodes can teach a Multitask DiT policy a new behavior, and the platform also runs Pi.05 and JEPA models.

Everything ships public

The robot is OSHWA certified (US002788), and the mechanical CAD, the PCBs, and the control software are published. The firmware lives in the Apache-2.0 cranebot3-firmware repo. The training data is the part other robot startups keep locked up: the naavox account on Hugging Face holds 141 public LeRobot-format datasets, grasping runs, laundry pickups, toy cleanups, visual servoing drills, including the bedroom-oct15 laundry set, with dataset names running as recent as October 21. There is also a controller-in-the-loop MuJoCo simulation, so you can develop against the kinematics before drilling into any studs.

For the agent crowd, the nf_robot Python API and an MCP server expose the robot's state to the same model context protocol your coding agents already speak. The pitch is a home-lab courier: ferry parts between workcells, hand a printed part from one robot arm to another, move laundry to a folding station.

Local by default

Stringman can run air-gapped: the UI is served locally, the AI models run on your own hardware, and no data leaves your network. A cloud telemetry relay exists for remote access through the playroom UI, it is opt-in, and you can drop back to LAN-only mode whenever you want. No subscription appears on the product pages.

Caveats

The comparison numbers are the maker's own, from idealized geometry: the arm comparison is calculated and ignores the arm's own link weight, and payload peaks near the room center where the lines share the load best. This is pioneer hardware by its own description, and the listing says so plainly: the autonomous behavior "may not always work," and buyers are expected to download updates and retrain models. Hacker News commenters flagged wobble in traverse video, lines crossing the room permanently while installed, and the need for strong, consistent indoor lighting for precision. Installation wants wall studs or structural columns, and a ceiling fixture hanging in the center of the room rules the room out. Build kits exist alongside the $795 assembled unit, and shipping currently covers the United States, Mexico, Japan, Korea, Australia, Vietnam, Taiwan and the UK.

Key Takeaways

• Neufangled Robotics sells Stringman, a $795 assembled open-source cable robot: four ceiling-mounted motors suspend a 2-finger gripper that moves up to 2 kg anywhere in rooms up to roughly 8 x 8 x 3 m.
• The maker's numbers: 4 actuators with 1.9 kg average payload versus 0.32 kg for a 7-actuator arm on the same motors, and about 500 times the work volume per dollar of a $2,500 research arm.
• The AI stack is local and open: a custom target-detection model plus LeRobot-trained grasping policies (Pi.05, JEPA, Multitask DiT), with about 50 recorded teleoperation episodes enough for DiT to learn a new behavior.
• Everything is published: OSHWA-certified hardware, Apache-2.0 firmware, PCB and CAD files, a MuJoCo sim, and 141 public LeRobot datasets on Hugging Face under naavox.
• An MCP server and the nf_robot Python API expose the robot to agents, pitched as a room-scale courier for home labs with multiple machines.
• Safety is a 28-newton line tension limit with damped motor release plus a hardware e-stop; the robot runs air-gapped with models on your own PC, and no subscription is required.

Sources: Neufangled Robotics (Stringman product pages), Stringman listing and specs, cranebot3-firmware (Apache-2.0), OSHWA certification US002788, naavox LeRobot datasets on Hugging Face, MCP server guide, Hacker News thread (619 points, maker answering), RuntimeWire on Neufangled and Nifong