Anomaly Detection. Beta testers wanted

ROS ROS General
1

Hi, I used to work at a robotics startup that built robotic dishwashers. (https://armstrong.ai/) And part of my job was to set up diagnostics, monitoring and alerting.

Aside from not being the most interesting part of my job, it was just plain tedious to manually set (and continually tune) thresholds for various signals, topics, etc. I frequently thought it would be nice to have a system that would automate that part of my job and now that the company has closed, I built it.

Now I have reached the point where the useful next step is real robots. The system performs well on recorded public datasets and on some production infrastructure (see below), but what I do not have yet is enough hours on a variety of real robots, doing real work.

Details are here at the Transitive Robotics website:

What it is

The system watches your robot’s topics and tells you when something is wrong. No threshold tuning, no hardcoded rules. It learns what normal looks like on your robot, then flags and reports anomalies in real time.

How to get involved

Reply to this thread or email shane@targetnode.ai. I will help you get it installed and monitored.

What you get and what I am asking for

What you get:

  • Anomaly detection for free on your robot with no manual threshold tuning to maintain.

  • Direct influence over the direction of the system/project. I want to know what is missing, what is confusing, and what breaks on real robots in production.

What I am asking for:

  • Honest and detailed feedback

Where it has been tested

I evaluated the system against public datasets that cover real hardware faults across very different platforms. The system has been tested successfully on the following datasets:

  • CASPER. A UR3e collaborative robot arm with real IMU fault data, from Cardiff University. (paper)

  • PADRE, Parrot Bebop 2. Real propeller damage on a quadrotor, recorded from onboard IMUs. (paper)

  • PADRE, 3DR Solo. Real propeller damage on a second quadrotor airframe, with IMU and barometer data.

  • ALFA. A CarbonZ T-28 fixed-wing aircraft with real engine and actuator faults, from Carnegie Mellon University’s AirLab. (data)

  • UAV-SEAD. A large set of real PX4 flights with naturally occurring state-estimation anomalies, from Istanbul Technical University. (paper)

Additionally, I am running the system on my production servers. They are not robots, per se, but they are live production systems that get used every day. Right now the system is watching CPU, memory, disk/network operations (29 signals total) all without me setting or tuning thresholds. In the last 200 hours of operation (since Tuesday, June 16, 2026) there have been no false alerts.

Thanks for checking out this post and if you have any questions, reach out to me at shane@targetnode.ai

4 Likes
ROS News for the Week of June 22nd, 2026
2

Detecting anomalies is such a universal need in robotics, especially when fleets grow from 5 to 50 robots and beyond. It’s also a really difficult problem to tackle, which is why I’m personally really excited about this new beta capability – and the first 3rd-party capability to pass our review. Curious to hear what people think after trying it out.

Would also love to hear what folks like @doisyg (Dexory), @Iliabara (Polymath), or @kiwibot-ai think of this! How do you currently do anomaly detection in the field – is it mostly hand-coded alerts or do you already use some kind of ML like this?

1 Like