Hack the North · built on BracketBot
We do the R&D,
so you can stay
on DND.
Some of us have woken up in the middle of the night, stumbled from room to room and lost hours resetting 3D printers. PrintDND clears the printer on its own and puts the print where you want it, so printing carries on while you stay on Do Not Disturb.
No teleoperation in that clip · the policy is driving
Introducing
PrintDND.
A robot that autonomously clears the printer from its previous job and places the print down in a desired location. That's the whole idea. It means printing can keep going while you're asleep, out, or on Do Not Disturb.
It isn't a script. The robot is running a learned policy, so when it misses a grasp it simply tries again.
Teleoperate,
then hand it over.
The robot came from BracketBot, a Hack the North sponsor: bimanual, 7-DOF per arm, on a two-wheel base that uses reinforcement learning to keep itself upright.
We took that stack and broke the job down from first principles — the rudimentary movements needed to get a print off a plate. Then we collected data by teleoperating the robot with a Meta Quest 3S, over 100 episodes, stored in BracketBot's cloud.
We audited that data for quality and used it to fine-tune Physical Intelligence's PI0.5 vision-language-action model, then built on the deployment stack to run inference. On top of the reset we wired up a wake word: say "Hey BracketBot, reset the printer" and it listens, answers, and goes.
- Platform
- BracketBot
- Arms
- bimanual, 7-DOF each
- Base
- 2 wheels, RL balance
- Teleoperation
- Meta Quest 3S
- Episodes collected
- 100+
- Policy
- PI0.5 VLA, fine-tuned
- Inference
- Montreal servers
- Voice
- "Hey BracketBot" wake word
- Printer
- Bambu Lab X1 Carbon
The robot alongside is replaying one of those recorded episodes — joint angles, gripper state and body lean are all logged values, not keyframes. Platform: bracketbot.com.
What broke.
Most of the weekend was spent here, so it would be strange not to say so.
-
01
The wifi
HTN, eduroam and even Starlink were spotty and inconsistent. We moved the entire setup three separate times, and each time the failure looked slightly different — pairing the Quest, uploading to the cloud, or just holding an SSH session.
-
02
Losing 75% of our data
There's a local
/pendingfolder that holds recordings until the bot can upload them. It has a cap, and past that it silently overwrites. We found that out far too late, after recording a long stretch with no wifi. -
03
A 2kg gripper and a stiff plate
The grippers top out around 2kg, well under what it takes to lift the magnetic build plate. So we printed a jig that flexes the plate instead and shakes the print loose — much less force, same result.
-
04
A blind head camera
The printer occludes the head camera almost perfectly, so the policy leans on the wrist cameras — and depth is hard to judge from one view. Early on, home position looked so much like the lift pose that the robot decided it had already finished before starting.
-
05
Speed
Planning every joint angle takes real compute. Running inference on Montreal servers took the edge off, but the robot still moves slowly. We're chasing something closer to home.
We envision a world
where robots build robots.
3D printing is where we started. It isn't where this stops.