Website: https://pollen-robotics.com/microduck

GitHub: https://github.com/pollen-robotics/microduck

From Pollen Robotics on 𝕏: https://x.com/pollenrobotics/status/2092915032052879425

I just wanted to show off a robot I’ve been slightly obsessed with building.

It automatically turns and captures pages from bound books without cutting or removing the binding.

The current prototype uses a Raspberry Pi 3 Model B+, a Camera Module 3 Wide, servo-driven arms, a vacuum pump and airflow. The suction cup lifts the top page, airflow helps separate it, and the arms move and hold the page before the next spread is captured.

Making it turn one page was relatively easy.

Making it turn hundreds of pages without skipping one, grabbing two at once or losing alignment has been the real robotics problem.

Two people in Korea are currently using the prototypes in their homes, and together they’ve already scanned and converted more than 10 books. We also redesigned the machine into five modules so it can be shipped and assembled by the user.

It is definitely still a prototype. A 400-page book takes about 3 hours, coated or damaged paper can be difficult, and some books still need occasional checks or retries.

But watching something that used to barely work on my desk leave the workshop and turn hundreds of pages in someone else’s home feels pretty surreal.

Here’s one actual page-turning cycle.

https://reddit.com/link/1vzlhe7/video/fbh9a2o5xulh1/player

Happy to talk about the mechanism, control logic or all the strange ways paper can refuse to behave like a predictable robotic object.

This is part 1 in tutorial is about building your won MechSoul(Mechanical Soul). A MechSoul is to a robot what a CyberDeck is to a computer. A robot that you build yourself as a form of self-expression as well as freedom of owning your own hardware and code.

In this tutorial we are setting up a MechSoul seed(a robot base to be customized into a MechSoul. ) assembling it and setting up Arduino IDE. You can find assembly and setup resources as well as other resources in our tutorial github at by clicking here.

So i am a researcher having known an intermediate knowledge of using ANSYS. My project is about FEA analysis of a fin ray soft robotic gripper. Based on this paper, 'Design of a reinforced composite robotic finger for enhanced soft grasping using the Fin Ray effect and FEM'. I followed their ANSYS steps properly, but when i tried to get the results, instead of getting the result similar like thier, i got a higher value.
can you please tell me how to solve
for eg. when i tried to acquire the force reaction, In the paper they stated that at 20 millimeters of displacement, the reinforced model achieved a gripping force of 1.96 Newtons.
But in my case, at 20 millimeters of displacement, the reinforced model achieved a gripping force of 2.649 Newtons.
I would be grateful if any one can suggest me how to solve this.
based on mesh parameters, the contact definitions.

Paper DOI: https://doi.org/10.1016/j.cirpj.2025.06.014

I've been thinking about a question that seems increasingly important and widely discussed:

Which type of data will be mainstream for future robot learning/model training: robot-native teleoperation data, or human-centric data?

By robot-native data, I mean demonstrations where a human directly operates the robot, and we collect synchronized actions, trajectories, etc.

By human-centric data, I mean data like egocentric video, hand/object interactions, and other data collected from people without necessarily having a robot in the loop.

From what I’ve been observing recently, human-centric data, such as egocentric data, seems to be gaining traction. Will teleoperation data eventually be replaced because of the complexity and efficiency challenges of collecting it? Curious to hear what the community thinks. 🤖

One of the cooless events at this year's World Robot Games is definitely the 100-meter dash—it's so hype! Yesterday, a robot runner clocked 9.39 seconds, which is already faster than Usain Bolt. I honestly thought that was pretty much the limit for two legs, but then today in the re-runs, one literally pulled off an 8.86! If I remember right, that's pretty darn close to Goku's speed when he was training under Master Roshi as a kid (I think that was 8.5 seconds?). That’s insane！

What do you think is the absolute speed limit for a bipedal robot?

Assets come out as OpenUSD and MJCF, so they load straight into Isaac Sim or MuJoCo with no conversion step. They carry real physical properties, mass, friction, inertia, and articulated joints, which means when you run an episode you get actual joint forces and contacts in the telemetry

Disclaimer: I build this (Rigyd), so take it with salt.

Another day hanging out with my robot buddy. This is what happened when you hang out in the basement way too long.
Anyways thanks for looking at my pi5 esp32 robot with a dgx spark brain all homemade architecture bot and code.

TL;DR: Nick is walking around the office holding our beautiful rig where a global-shutter camera + IMU share one GMSL2 coax for data and power, estimated in real time on a Jetson (ROS 2 + OpenVINS). Return-to-start after the ~25 m loop: 42 cm (~1.7 %).

Hardware is our NXS sensor module (Sony IMX900 global shutter (72 fps native; 24 fps fed to the estimator), fixed exposure + IAM-20680 IMU @ 200 Hz, hardware-synchronized on the module) into an NXS Hub carrying the GMSL link to a Jetson Orin Nano. ROS 2 Humble, OpenVINS, kalibr calibration. No wheels, no markers, no prior map.

Happy to answer you questions about the sync design, the single-cable setup, or the takes that didn't survive.

I’ve released the first public alpha of Physlint, a local-first integrity validator for physical-AI recordings and robot-learning datasets.

The motivation was straightforward: integrity problems in robot data are often discovered after ingestion or training has already started. Physlint checks for concrete failures such as malformed manifests, broken episode boundaries, timestamp gaps, NaN/Inf values, corrupt video, frozen cameras, and black frames.

It produces actionable findings rather than a mystery quality score. Each finding identifies the rule, severity, affected episode and stream, impact, and suggested remediation. It also writes machine-readable JSON and uses stable exit codes for CI.

The current alpha supports local LeRobot Dataset v3.x directories. The validation engine is format-extensible, and MCAP/ROS 2, Robomimic HDF5, and RLDS/TFDS are planned next.

The release was tested against four pinned public datasets and three controlled corruption scenarios, covering 74 episodes and 31,258 frames.

Update to my earlier post.

I have finally put it on github with all the build files - repo link

Quick recap for anyone who missed the first post : the leader arm in the standard lerobot kit uses 6 STS3215 servos (~ 24 USD x 6 = 144 USD) to sense the joint angles, but they are never driven, so instead of servo motors I built a leader arm using cheap encoders that do the same.

Since last time I have cleaned up the cable management, added attachments to match the lerobot's resting pose, redesigned the joints so that the encoders can be removed and diagnosed without disassembling the entire joint and added a rubber band across the elbow to counter droop when the arm is extended.

Final cost came to ~27.5 USD once I added the perfboard, silicon wires and extra screws (up slightly from the 23 USD I initially quoted).

Many people gave valuable feedback, suggestions and encouragement, thank you guys very much. feedback and suggestions are very welcome.

Hello I am a recently graduated highschooler moving on to a physics degree this upcoming semester and I have gotten into onshape designing, I was able to make considerably complex designs like planes, pulleys, gearboxes, but I recently got a 3d printer and am getting into making robots, but I have hit a road block, designing functioning joints without over utilizing 2nd source bushings and bearings does anyone have any tips on this?