top of page

semarx

Robotic Arm Mechanism

Robot wear detection with zero training data

Detects friction, wear and added load directly from the control-loop signals your robot already produces. No training, no robot model, under one millisecond per cycle. Blind test: you send logs with faults you already know about, unlabeled; we return what we detect and when; you score us against your own records, before any commitment.

Robotic Arm Mechanism
Comparison for robots: usual monitoring reads one signal summary or a trained model; the robot information monitor checks whether command, response and the next response still fit together, needs only 360 averages from normal cycles, and gives a first hint of which side changed.
Model size held at run time: Semarx loop monitor uses 360 values with nothing trained, versus 1,015,560 neural-network weights for the best learned method (MVT-Flow)

Results on real robot data

Tested on real robot data: a physical six-axis collaborative arm, 2,122 recordings and 12 hardware-induced fault types. Ahead of six of seven published baselines and within 5 AUROC points of the best, while retaining about 2,800× fewer values and no learned weights.

Chart: mean AUROC on real robot data across 12 fault types. Semarx loop monitor scores 88.6 with no learned weights, versus seven published baselines from 77.5 to 93.6
Semarx loop monitor vs MVT-Flow: 93.0 vs 84.7 AUROC on collision with cables, 91.0 vs 87.5 on collision with foam, and a more consistent spread across fault types. The anomaly reading rises with added load on the arm

How it works

Your controller already records what the robot sensed, what it commanded, and what happened next — every cycle. We read how those three stay related. When friction, wear or an added load changes that relationship, the reading moves, even when position tracking still looks perfect.

Comparison for robots: usual monitoring reads one signal summary or a trained model; the robot information monitor checks whether command, response and the next response still fit together, needs only 360 averages from normal cycles, and gives a first hint of which side changed.

Test it on your own data

  1. Send us logs. Normal cycles, plus a log containing faults you know about but don't label.

  2. We analyse blind. We return the timestamps and severity of what we detect.

  3. You score us. You hold the answers; we see them only after we submit ours.

​

​

Is this for you?

Industrial robot arms, collaborative robots, pick-and-place cells, servo axes and motion stages, mobile manipulators

​These are examples; any controlled machine that logs its commands and measurements can be tested. A robot with a learned controller: choose this page for hardware wear, RL Agent Monitoring for policy drift.

 

What to send: two files, a normal baseline (ideally about 1,000 normal cycles; fewer is fine to start) and a test log with faults you know about, labels removed. Core columns per joint: measured and commanded position, and motor current, with a timestamp and cycle ID.

Industrial robot arms, collaborative robots, pick-and-place cells, servo axes and motion stages, mobile manipulators
bottom of page