Stringman: A $795 Room-Scale Cable Robot for Picking Up Toys

Background and Phenomenon: A Toy-Picking Machine Hanging from the Roof’s Corners

On November 3, 2026, Hacker News’ Show HN section featured a robot named Stringman, produced by Neufangled Robotics, priced at $795. During verification, there were nearly 300▲ and 115 reviews. Its shape defied existing categories: no wheeled chassis, no leg joints, four cables hanging from the roof’s corners, with a gripper suspended in mid-air, moving around the room via motors that extended and retracted the cables. The official website’s slogan borrowed a sentence from the Transformer paper: “Tension is all you need.”

The second breakdown point was open-source attitude: the website listed “Open source and air gappable” as a selling point on product pages, certified by OSHWA (Open Source Hardware Association), with CAD, PCB, and control software all publicly available. It supported operation over a local network, with the UI running locally and AI models running on user hardware, ensuring no data went online. Cloud relays were optional, and the official statement stated that users’ videos were not stored and they could always return to local network mode.

The repository matrix (verified snapshot on November 4): The control code nhnifong/cranebot3-firmware had 133 stars, under Apache-2.0 license, created on January 14, 2025, with the last push on November 2, 2026. Code was still being pushed the night before HN’s release. The PCB repository stringman-pcbs had only 3 stars and no README. The gripper arm’s source was clearly indicated: based on TheRobotStudio/SO-ARM100 (7979 stars), part of the LeRobot ecosystem.

Architecture and Technical Principles: The number of actuators determines the price, so only four are used

The official website directly stated the industry rule: “In robotics, price is basically actuator count.” Stringman’s solution was to use 4 motors to cover the entire room: a parallel cable mechanism, with lifting forces shared by multiple motors, and load distributed according to joint torque rather than individual actuators, achieving the most even distribution near the room center.

The MCP guide in the repository provided a more precise structure than the official website: four cables, one tension point at each corner of the roof, with only the diagonal two corners being electric anchor points (each containing two motors and one camera), and the other two corners being passive cable-eye rings. The expensive motors and drive boards were concentrated in the two corner units, while the other two corners had near-zero cost.

Specifications (official website’s own report): The working area covers rooms with diagonals of 3 to 10 meters and ceiling heights of 2.3 to 4 meters, equivalent to 147 cubic meters at a size of 7×7×3 meters. Average load is 1.9kg, reaching up to 2.2kg near the ground. The maximum tension of 28N minus the weight of 638g of the gripper. Side movement speed is about 0.4m/s. Comparison table on the official website: A 0.75m series-connected arm using DM-H6215 actuators (7 actuators) has an average load of only 0.32kg; compared to the $2500 SO-101 arm, Stringman’s cost per cubic meter of working area is 185 cubic meters versus 0.35 cubic meters, about 500 times higher (official website’s own report, ignoring the weight of connecting rods for the series-connected arm).

The BOM list for the hardware directory is even more extreme: four DM-H6215 brushless hub motors cost $66 each, totaling $264, making it the most expensive item. The cable body is a 40-meter, 50-pound braided fishing line, priced at $5.83; the cable guide ring is an Alconite guide ring used for fishing rods, costing $0.91; slip rings, force sensors, laser rangefinders, and carbon fiber suspension rods all cost no more than $7. The total value of “total per kit” in the BOM list is approximately $646 (aggregated based on small batch purchase prices), which means a very thin profit margin compared to the $795 selling price. The entire device is 3D printed (structural parts made of PET-GF, decorative parts from PLA); it operates off mains power without batteries, with the gripper powered by a FEP dual-core cable that can withstand 50N, eliminating the need for a battery and downtime during charging.

There are two notable design features in terms of mechanics. First, swing elimination: the gripper is mounted below the pivot point, allowing it to fit under the bed or table, but at the cost of swaying like a pendulum. Stringman does not use motors but instead uses an internal IMU in the gripper to implement an active swing elimination algorithm, trading computational power for hardware. Second, passive safety: the maximum tension of the cable is 28N, roughly the weight of a backpack. When the timer is reached, all motors relax, allowing the cable to be directly pulled off; most of the time, the cable is gathered near the roof. Installation requires only a hand drill and a Ladder Platform, with calibration taking about 5 minutes without physical intervention. The anchor points are located below the fan blades, so the cable cannot reach the ceiling fan.

Key data: 66,872 lines of Python, 51 test files, and one MCP bridge

Local read-only clone tested (verified snapshot on 11-04): 433 files, 66,872 lines of Python, consistent with GitHub language tagging. Top-level directory: src/nf_robot (control body), simulator (MuJoCo simulation), playroom-ui (frontend), hardware (BOM and circuit board), tests, stringman-pilot-rpi-image (self-made Raspberry Pi image), along with experiments, scripts, and PyInstaller packaged files. The pyproject version is already 7.5.0, but the repository name is still cranebot3. The software and hardware versions follow separate timelines, a common issue in open-source hardware projects.

Distribution does not require a clone repository: pip install "nf_robot[host]". The main entry point for the motion controller stringman-headless is observer.py. The README describes it as the central brain of a single robot, communicating with the anchor points and grippers via WiFi, listening on port 4245, sending and receiving protobuf messages, with the local UI running at localhost:8090. PyTorch and CUDA dependencies amount to about 5.5GB, with all computing power external: the user’s PC serves as the controller (minimum 8 cores, 8GB memory, recommended for miniPCs or Macs with NPU).

There are three aspects in the AI side. First, it is compatible with the LeRobot control stack: it supports policies such as Pi.05, JEPA, Multitask DiT, etc. Training data can be collected in the UI, and DiT models require at least 50 episodes to learn a new behavior (as reported on the official website). The dataset is publicly available on Hugging Face. Second, semantic fiducial tags: clip-style labels are used to mark destinations like laundry baskets, toy boxes, and trash cans, without needing to label individual items. The official website states that the network has learned “toybox” to refer to a specific tag, enabling it to identify containers beyond the appearance of the training set and serving as both source and destination for delivery tasks. The official website describes it as “the courier of home laboratories.” Third, the MCP server: the interface is explicitly designed for agents.

The MCP bridge is the most worth-decomposing part of the repository. It is a module under host (mcp_bridge) that runs within the observer process, allowing agents to hot-load new maneuver actions without restarting the controller. Design constraints are written in code comments: the bridge itself is just an ordinary maneuver, with all actions performed through the observer’s public methods. The stop button, tension monitor, and parking status check apply equally to agents. GUIDE.md sets even stricter rules for agents: before the first movement within a conversation, the user must be asked if it is safe; stop is always effective, equivalent to a UI stop button; no weak security monitor should be adjusted without permission; agents should not answer physical confirmation pop-ups for users; shutdown and firmware updates are directly refused, leaving it to user operations in the UI.

The core of the perception side is an “ortho map”: two anchor camera views combined into a top-down projection of the ground (5 meters correspond to 512 pixels), with agents and policies viewing the same image. There are 51 tests in total, with specialized tests for decluttering, derailment detection, and MCP bridge. The simulator uses MuJoCo instead of real hardware, but runs real component firmware service processes, which is controller-in-the-loop simulation.

Impact and Significance: The mechanism selection re-negotiated prices, and a security manual was written for agents

First, the cost-reducing approach of “mechanism priority.” The working volume of the series-arm increases with the cube of arm length, making actuators more expensive. The cable parallel mechanism outsources lifting to the room structure, saving 147 cubic meters with four motors, achieving about 500 times the working volume per thousand dollars (as reported on the official website). The cost is mobility loss: it cannot move outside the room, and installation requires drilling holes or binding structural parts. The official website’s FAQ is also straightforward: it cannot be installed if there are low-hanging obstacles in the middle of the room.

Second, agent interfaces represent a risky investment. Writing the MCP server with consumer-grade hardware is rare, and writing top-tier security documentation for AI agents that includes “ask the user first, stop is always available, no weak monitor adjustment, no answering pop-ups for users” is even rarer. Agent permissions are confined to the same set of security monitors as UI buttons, with no bypass channels at the interface level. This is isomorphic to the software agent field’s concept of “the host holds the permissions, messages are just inputs.”

Third, pressure test samples for the open-source hardware business model: full design disclosure, Apache-2.0 license, with upstream development supported by SO-ARM100, which has 7979 stars. It sells complete machines, cloud relays, and maintains an account system. The implicit argument is to use openness as a moat: BOMs are disclosed, profits are low, and copying profits are limited; the community fixes bugs, while the official side earns money from selling machines and services.

Fourth, educational and research focus: inverse kinematics, telemetry, and imitation learning can all be taught on this “flying gripper”; cross-formulation is a major challenge in robotics, driving completely different body configurations, which is a stronger outcome than transferring between identical双臂. Public datasets, simulation, and $795 hardware make it the cheapest experimental platform in this direction.

Controversies and Risks: The cable crossing living space is an issue that cannot be fully debated

First, safety controversies are the most vocal in the comments section. One highly praised comment means: everything was great until something suddenly pulled on the cable. Another thinks picking up clothes seems silly, but picking up toys is truly amazing, though the cable may never work in a household environment. It’s more interesting to know where it will fall in commercial scenarios. The 28N relaxation mechanism is honest passive safety, but the official FAQ also acknowledges: as long as it’s still in use, the cable will cross the room, and the storage state is close to the ceiling. What does having a cable crossing living space for a long time mean for families with children and pets? No one has given a reassuring answer yet (this is my judgment).

Second, there is tension between open-source attitude and commercialization path. On one side, it’s air-gapped, LAN-only, without storing videos; on the other side, there are account binding, management consoles, telemetry relays, component firmware images downloaded from official storage, and MCP layer to control machine and firmware updates left to the official UI. The official rhetoric is that privacy and convenience are choices, and the decision lies with you, not us; but the default software and firmware channels are in the hands of the official. Component Raspberry ships with a unified default password, and the README teaches users to deploy SSH keys and log out with passwords. This is frankly commendable, but most household users won’t go this step (this is my judgment).

Third, community enthusiasm should be read calmly. The HN has nearly 300▲ and 115 comments, with 133 stars on GitHub (verified snapshot on 11-04), about 60 times more than upstream SO-ARM100’s 7979 stars. Interest is on HN and the official website, not on GitHub, and star ratings have not yet translated into community contributions. Some people in the comments are worried that big companies will copy the concept and then refine it. Responders remind that cable robots have existed for decades, and those who wanted to copy them have already done so; others recall how they spent an hour every day picking up toys thrown by their younger siblings when they were kids. Having a machine like this could save hundreds of hours.

Fourth, the boundaries of numbers. 1.9kg load, about 500 times, about 1000 hours of real household testing, 50 episodes of learned behavior—all reported on the official website, with no third-party retesting yet. The 500 times is based on “every thousand dollars of working volume,” but changing it to every thousand dollars of effective load or number of tasks completed would lead to completely different conclusions (this is my judgment). There are also more specific questions in the comments: how to handle dirty scenarios in real households like Lego with chocolate milk spilled, cereal bowls pressed on toys, and blankets covering toys.

There’s also a separate debate in the comments section: Some believe that the appeal of humanoid robots is purely aesthetic and almost inevitably inferior, with hanging from ceilings being a more plausible approach, followed by R2D2-style rolling chassis. Others use black humor to mock the consequences of humanoid robots recognizing their host as an inefficient consumption of corporate resources. The response is: What’s needed is the most efficient design for completing tasks, even if it looks like a hybrid of an octopus and horror movie characters.

Conclusion and Outlook: Low cost itself is a technical approach

The core argument can be summed up in one sentence: If the actuators are expensive, then use room structure as free actuator mounting points, fishing line as free transmission components, and algorithms as free balancing weights. Four motors can create a $795 robot with room dimensions. The software framework (66,872 lines of Python, 51 tests, controller in-sim simulation, hot-loaded MCP bridge, homemade Raspberry Pi image) shows this isn’t a maker toy but a machine engineered according to product standards.

Two things worth following up on: Whether the open-source and low-profit business model can survive, and whether the 133-star and 3-star PCB repositories indicate that community contributions have not yet begun. Also, the cross-formulation progress of the LeRobot ecosystem—whether models can be migrated to other body configurations will be a harder test than the official website descriptions.

My conclusion in one sentence: Stringman is worth analyzing not because it predicts the ultimate form of home robots, but because it proves that when mechanism selection redefines price, software frameworks, agent interfaces, and open-source attitude will all be redefined.

Reference Sources

• Official product page and FAQ: https://neufangled.com (accessed on 2026-11-04; 1.9kg, about 500 times, about 1000 hours, 50 episodes, etc., all self-reported by the official site)
• HN Show HN: story 50354358 (posted on 2026-11-03, with nearly 300▲ / 115 comments during verification)
• Control repository: https://github.com/nhnifong/cranebot3-firmware (133★ / Apache-2.0 / created on 2025-01-14 / pushed on 2026-11-02, verified snapshot on 11-04)
• PCB repository: https://github.com/nhnifong/stringman-pcbs (3★, verified snapshot on 11-04)
• Upper-stream gripper arm: https://github.com/TheRobotStudio/SO-ARM100 (7979★, verified snapshot on 11-04)
• Local read-only clone verification (2026-11-04, no code running in any repository): README.md, pyproject.toml, hardware/bill_of_materials.csv, src/nf_robot/host/ (observer.py, mcp_bridge/), simulator/README.md, tests/, stringman-pilot-rpi-image/
• HN comment pool: hn-stringman-comments.txt (all text is quoted, not directly cited)

Author: Yongliang