Stringman Production and Feature Update
Nathaniel Nifong
Stringman can currently be pre-ordered for a refundable $120 deposit. The full price is $965 with the difference paid at shipping time.
The reason for doing this is that one of the core components, the Raspberry Pi Zero 2W, is experiencing a huge supply shock due to AI memory demand.
According to Raspberry Pi, they are expected to become available again in January 2027 which is four months later than I had guessed. Stringman robots are otherwise ready. So meanwhile I’m designing a variant that uses the Pi 3 A+ which is readily available. I will send updates as soon as possible on whether the shipping timeline can be accelerated with this substitution. It’s likely fine but will make the gripper look bulkier until we can switch back to the smaller pi.
We have over 1400 episodes of collecting and putting away items with stringman, and a recipe based tool for deriving datasets from it.
The data exists at full resolution in a universal format which more specific formats are derived from in order to test all kinds of different models and approaches quickly.
Any model type trained for Stringman can be run by just entering it’s name in the policy box.
naavox/dit-grasp-4 - grasps any object in view of the gripper with a diffusion transformer
naavox/naavox/xvla-move-clutter-3 - uses XVLA to find and put away objects using velocity control
naavox/xvla-camera-goal-waypoints-2 - uses XVLA to control the robot by predicting waypoints
naavox/pi-move-1 - is a fine-tune of Pi0.5 from Physical Intelligence to find and put away objects.
The reason for trying all these methods is that Stringman is so different from a bi-manual arm setup that it’s still anyone’s guess what the best approach is for controlling it with AI. If this kind of experimentation is something that interests you, please get in touch.
It might seem like other topics deserve development, but all other topics are basically calibration topics when you get into the details.
High speed and smooth motion depends on accurate self calibration. Flexible dataset tooling, good model results, and low latency teleoperation are all downstream of accurate calibration. When Stringman is installed in a new room, it spends about 4 minutes performing specific motions to measure its components, sensors, and cameras, and the result of this determines everything about the quality and reliability of other behavior. In the past month I have put a tremendous amount of testing into calibration and made several improvements.
Calibration is now three stage, witch each stage refining the result of the prior one. If any stage regresses, the result isn’t used. This prevents many of the “tension overload” events people encountered during calibration.
The new third calibration stage makes measurements that are used for a new class of residuals that better constrain the optimization problem.
Automatic swing cancellation tuning - A major step in calibration is now to automatically tune the swing cancellation to the specific network and machine environment.
Safety checks - At multiple stages calibration automatically checks for and prevents the use of degenerate results so Stringman never makes unexpected bad moves.
Flexible adaptation to various room sizes.
Hardware improvements from early adopter feedback
The true test of a product is to ship it. And when the first Stringman robots were shipped, parts of the decorative anchor covers broke in transit. Just calling them decorative was half of the problem. In shopping they bear the weight of the main motors so they’re actually structural during that time and needed to have sufficient thickness and ribbing to handle the loads they experience. These parts as well as others have been beefed up.
The big box has been swapped for a carbon fiber pole and flat marker. It’s hard to eliminate the marker from the design but it is less intrusive now and much stronger. All pole joints are made with epoxy now instead of super glue.
Multiple customers encountered poor uptime from anchors and the ultimate cause is overheating. Bigger heatsinks, better placements, and more airflow were tested, with airflow being the only thing that helps enough to solve the problem. All Stringman robots will be shipped from now on with more powerful cooling fans.
Optimizations, Ease of use, and Privacy
There have been 230 commits since my last email. Nearly the whole codebase has turned over. The biggest improvement is that CPU core use has been further slashed by finding places where compressed video can remain compressed and merely re-muxed on its way to its destination. This lowers latency and makes it easier for humans and AIs to control stringman. Check out the details if you’re interested in the System Architecture of Stringman
All UI elements have tooltips to help with learning what the buttons do, and there’s an interactive tutorial available at neufangled.com/tutorial.
The UI itself has been completely factored out of neufangled.com and moved into a standalone NPM package, and is now hosted locally by the robot motion controller. This affords two things: one, the system is truly air-gappable, and two, users who fork cranebot3-firmware can update everything with their agent and no have to clone two repos.
Crowd Supply Demo
Neufangled Robotics presented Stringman at Crowd Supply 2026 where attendees could test drive it along with dozens of other inventions and robots.
.jpg?part=0.2&view=1)
As always I am looking for anyone in NC who would be willing to let me come install Stringman in their home and collect training data. If that's something you want to try let me know.