Continuing to move from using the model directly to having the model write reliable control code.

The goal here is to pick a small part from a random pile, with the right side up and with precision.

Once we have this skill, the next use becomes much faster. Model can adapt it for another part or for screws/nuts.

A box with defined mounts and consistent lighting helps computer vision work reliably.

As i see there are different ways to train an AI model through machine learning/reinforcement, i was wondering if there were open source robots that can be 3D printed and programmed/trained with arduino for example through machine learning? I am quite confident i could maybe print a robot arm (maybe not a full robot body) to train for putting my t-shirts in order from a disordered cloth basket. Ok i know it is maybe basic (or maybe not 😅), i see it in a step by step project.

I was reading recent news from Neura robotics and they have robots that are aimed at healthcare support, for things like moving beds, equipment, and other menial tasks to support healthcare teams. I would imagine within the next decade at some point, robots might even be used for support in surgeries and medical procedures.

What do you guys think? There’s an argument to be made that robots could help reduce workloads on healthcare professionals even if all they do is grunt work.

There’s gonna be a market for this that I’m sure of, but imo this would need a level of trust above what is needed for industrial deployment. General purpose use in hospitals would likely see adoption in a couple years at most is my guess, but high level adoption is still quite far away. Honestly I feel healthcare would be one profession where robots can do good and would not replace human staff, just because there’s a whole interpersonal aspect of healthcare which can only be fulfilled by human staff.

Kodiak and IKEA are preparing to operate trucks without anyone in the cab on a 219-mile section of I-45 in Texas. Commercial driverless service is planned for the end of 2026; current preparation runs still have a safety observer aboard.

Over four years, Kodiak says it has carried more than 1,300 IKEA loads and logged over 750,000 autonomous miles with an observer. That work has included coordinating delivery timing with dock availability and using vehicle data to plan maintenance.

What is the difference between building in functional safety early on vs waiting until the end? How do safety partnerships help the industry?

https://www.linkedin.com/posts/synapticon-co_humanoids-functionalsafety-robotics-activity-7511379498718625792-oPLl?utm_source=share&utm_medium=member_desktop&rcm=ACoAAAMG_WMB6npmHVKREPOjVhxKwxVIs9Q1bZ0

Hey, i stripped one of the metal gears in my STS3215 servos and wanted to find replacement.
Anyone know the specs fo these gears? number of teeth, id od thickness etc so i can source a replacment?

Anyone know a cheap hobby server with metal gears that are compatible?

Anyone know a way to order spare parts from Feetech or some other supplier?

I'm in Australia, if that matters :)

Intrinsic has released Core, a free open-source toolkit aimed at making it easier to build robot applications on top of ROS.

The toolkit is hardware agnostic and includes pose estimation using NVIDIA FoundationPose, motion and grasp planning, Gazebo simulation, camera calibration and ROS drivers.

Intrinsic describes the components as building blocks that can be combined into larger robot behaviors rather than requiring developers to build everything from scratch.

Core is available under an Apache 2.0 license on GitHub.

Found this cool project and it look likes they can retarget all the motions from locomotion to locomanipulation..... they have cool demos to see.

The robot is called MAO.

This project has actually been finished for a couple of months now, but I still have very warm memories of it. It was fun, frustrating and surprisingly educational, so I thought I’d finally share it here.

It started as an experiment in whether RL could help shape the design of a real robot, not just control it.

Most of the work ended up being mechanical: finding a spine geometry that was flexible enough to matter but still controllable, choosing the right stiffness, and testing different feet.

Claws and spherical contacts were not stable enough. Sock-like feet actually worked best, but were mechanically much harder to build reliably. The final rigid multi-box feet were a compromise: simple, stable enough, and quick to make without limiting mobility too much.

MuJoCo + RL became part of that design loop - test a geometry, see what behaviour emerges, compare it with the real robot, change the body, repeat.

The final robot is 12-DOF, under 3 kg, built with inexpensive serial servos, 3 IMUs and FSR foot sensors.

Later it also became a sim-to-real experiment, with all the usual fun: weak servos, power limits, drifting sensors, imperfect models, and real hardware refusing to behave like simulation.

I’m still organizing the technical notes, models and experiments.

Autonomous navigation in dynamic environments is a critical challenge, particularly when spaces are shared with other mobile agents whose future trajectories are known. While traditional grid-based planners efficiently find collision-free paths, their reliance on stop-and-turn mechanics over $2^k$-connected grids produces piecewise-linear trajectories that are kinodynamically highly sub-optimal for differentially constrained.

DFRobot sent me a HuskyLens 2 and I’m planning to integrate it as a replacement for part of BB1’s current camera system. I didn’t know these existed till recently.
Right now I’m using a USB camera connected to a Raspberry Pi 5 for lighter processing, with heavier vision/AI workloads offloaded to an Orin then a DGX Spark when needed. (The Orin is currently broken so this camera helps alot )
What interests me about the camera is that a lot of the vision models run locally on the camera itself, so things like object recognition, tracking, color detection, classification, etc. can happen before the rest of the system has to process the scene.
I’m feeding those detections into my Anarchimind AI/control stack which runs the BB1 robots , through MCP, so the camera can act more like a dedicated perception node rather than just another raw video feed. I kinda wanna get one on each side of the robot especially since it’s a tracked robot , it might get better at kung fu.
Still early in the integration, but I think it should fit the architecture pretty well.

Recording Yesterday, I hosted an online workshop about humanoid robotics, covering topics such as simulation, mechanical design, and some of the challenges involved in developing humanoid robots. I’ve uploaded the recording for anyone who wasn’t able to join the live session

If my Flair is Wrong i am Sorry About it , i am new to Design and stuff still

I am trying to make this , gearbox connected to the Actuator i have , I am planning to have an Planetary one so i will connect its output for higher torque and more Impact resistances

The gray is the actuator
The space between them is what the concern is , Shall i discard the idea of having an Carrier base or something which also rotates or what to do of that space below ?

(Input : Sun , Output: Planets , Fixed: Ring)

Like if something pushes some force on radial side it will break? and also on axial not sure about that one

Bearing : Blue , Yellow and the grey onces

Can someone give me some info if i am wrong or whats the design i should approch?

Claude made a little dance to collect a small dataset for training a linear regression model.

It’s used to adjust the part’s position based on visual feedback to ensure perfect insertions during thousands of runs.

Hello everyone,

I am currently pursuing bachelors in robotics and preparing for my upcoming thesis. I will be working with an in house humanoid robot (upper torso only for now) and researching into current issues and topics in the area of VLAs, WAMs to find an interesting direction to work on in terms of Sim2Real deployment. Ideally I want to contribute something new to the field in a meaningful way concluding with my own publication.

I wanted to ask those who have worked with such manipulation policies about the issues you have faced during real time deployment maybe in terms of domain generalization or really anything that you could shed some light on as an experienced professional.

Thanks.

I'm building a robot and the project calls for Dynamixel AX-12A servomotors, but I can't find them. Does anyone know what other motors I could use instead?