Yunho Kim

We present "hybrid system" that supplements conventional automation with "learning" for task & safety-level adaptiveness Deployed in factory for motor cable soldering (< 0.6 mm tolerance), resulting 108 motors, 99.4% SR with < 20 min data per task Paper: arxiv.org/abs/2604.22235

Robots that actually WORK in real factories. Without endless retraining. > Just 20 minutes of data. > 99.4% success rate. > 108 motors soldered in 5+ hours straight. Sub-0.6 mm precision on messy, deformable (!) cables. This hybrid “learning-augmented” system adds neural brains + 3D safety monitoring to ordinary cobots… and suddenly complex factory tasks become reliable, human-safe, and stupidly fast. This is for all of you who are into manufacturing. Huge credit to robotics engineer Yunho Kim @awesomericky99 📍Paper: arxiv.org/abs/2604.22235

@vai_viswanathan The best eval we did is checking whether it works well in real world deployment while matching QC. Other than that, there are additional experiments in the paper.

@AnneliesGamble @awesomericky99 @junja941 how do you eval these systems?

E2E learned robot policies aren't ready to be deployed in mfg (despite what a lot of people are saying rn). But a hybrid architecture like what @awesomericky99 @junja941 and their team built is. They're using conventional automation as the backbone, and adding in learning only where it's needed @Ken_Goldberg and I talked about a similar pattern a few weeks ago. The deployable systems are hybrid: deterministic backbone, learning only where adaptation is required, GOFE everywhere else. x.com/AnneliesGamble…

@theotime_lh Thank you Theo!!

@awesomericky99 Wow impressive! Nice work Yunho ! Industrial applications are the good target!

We present "hybrid system" that supplements conventional automation with "learning" for task & safety-level adaptiveness Deployed in factory for motor cable soldering (< 0.6 mm tolerance), resulting 108 motors, 99.4% SR with < 20 min data per task Paper: arxiv.org/abs/2604.22235

@nurvai_ai I am not fully sure about the mentioned drift. Learned controllers are modules inside a program tree. If success, it moves to next. If not, it retries.

@awesomericky99 Blending automation with learning for tight tolerance tasks makes sense, especially for reliability. Impressive results with so little data. How do you decide what stays rule based versus learned, and how do you handle drift in the learned part?

@nurvai_ai We use learning where variations are inevitable (e.g., cable bending). While current automation often uses hardware-centric solutions that can be costly and increase physical footprint, we instead take a software-centric approach to handle these variations.

In the end, it was not about fancy large models or massive datasets. We chose to harness small learned models within well-established workflow tools such as state machines and behavior trees. That made the system more predictable, debuggable, and easier to integrate.

@merelnyc With NN controllers/safety monitors, robot handles variations during the automation (e.g., cable bending, hole position, motor pose, nearby human workers ...). The type of task that robot handles are fixed as the overall task procedure is designed to accomplish motor soldering.

@awesomericky99 Impressive precision! Curious, how does it handle unexpected task variations? 🤔

We’re seeing a new wave of hybrid systems where conventional automation is enhanced with learning for both task execution and safety-level adaptiveness. In a recent factory application for motor cable soldering (<0.6 mm tolerance), this approach achieved 99.4% success rate across 100+ units, using less than 20 minutes of data per task. Faster deployment. Higher precision. Smarter automation. 🎥 Source: Yunho Kim ( @awesomericky99 - X ) ⚠️ This content is shared for informational purposes only. CTO Robotics Media is a media platform and does not own or develop the technology shown. Credit belongs to the original creators. #Robotics #Automation #AI #SmartFactory #Industry40

@vikkeycodes They are cobots. They have harmonic actuators

@awesomericky99 What kind of actuators does this robot use?

Here we show one of our experiments on AI + automation in factory. Built from certified industrial hardware, the system is fully compliant in the industrial enviroment. It achieved: 1. 99.4% SR (verified by actual QC tests) 2. 5.5 hours continuous run 3. Near-human takt time

Scientists have developed a control framework that enables #Raibo, a legged robot, to perform #parkour-like maneuvers and navigate complex environments autonomously. @kaistpr @HwangboJemin Read more in Science #Robotics: scim.ag/4dw3tAE

Our team is reaching out this year, sharing our research results and looking for excellent roboticists in Seoul. ai.neuromeka.com

Full marathone run in 4 hours and 19 minutes! youtube.com/watch?v=P8Tu_B…

Some part of our ongoing work on the news :). Just sprinkled some RL into it

Our paper is now published in IEEE Robotics and Automation Letters (RA-L)! Paper: ieeexplore.ieee.org/document/10705… It was a fun project to use egocentric videos of pedestrians, which can be obtained in large quantities, for autonomous mobile robot navigation.

A recent paper from researchers @KAISTCoE illustrates how to simplify the training of reinforcement learning policies by proposing a novel approach that they apply to a variety of legged robotic systems. ieeexplore.ieee.org/document/10530… #LeggedRobots #NeuralNetworks #QuadrupedalRobots

@ChongZitaZhang Thanks!!

@awesomericky99 So productive with solid works 🫡

Our work introduces an effective method to train a semantic traversability estimator with “egocentric videos” (video obtained by mounting a GoPro camera to a pedestrian's chest) and an “automated annotation strategy”. Paper: arxiv.org/abs/2406.02989 Video: youtu.be/EUVoH-wA-lA?si…

@catachiii 😀😁

@awesomericky99 You are on fire! 🔥

@servo_boyd Great to hear!

One of my favorite papers from this last year.