COFFAIL: A Dataset of Successful and Failed Robot Coffee-Making Skill Executions

Forum topic · 小凯 · 2026-04-21

Summary

COFFAIL (Coffee Failure) is a robotics dataset that records both successful and anomalous executions of a robot named Jessie performing seven coffee-preparation skills, including picking up cups, operating a coffee machine, adding milk, stirring, and delivering cups. Featured in a zhichai.net forum post, the dataset addresses a common gap in robot learning research: most existing datasets contain only successful demonstrations, leaving models unprepared for real-world anomalies. COFFAIL captures diverse failure types—grasping slips, placement offsets, wrong action order, and excessive or insufficient force—in real physical interactions rather than simulation, with complete episodes from start to finish. Failure data is critical for training anomaly detection models, developing recovery strategies, and evaluating robustness, since models trained only on successes suffer from distribution shift, fragility, and safety risks when deployed. The forum post compares the approach to aviation safety practices, where analyzing failures rather than repeating successes drives improvement. Paper: arXiv:2604.18126.

> *COFFAIL: A Dataset of Successful and Anomalous Robot Skill Executions in the Context of Coffee Preparation* > Alex Mitrevski, Ayush Salunke et al. | arXiv: 2604.18126 | 2026

An Overlooked Data Gap

Suppose you train a robot to make coffee. It learns to pick up a cup, press the coffee machine's button, wait for extraction, and hand you the cup—with a 95% success rate.

But what about the 5% of failures? A cup slipping, a button pressed in the wrong spot, spilled coffee, a handle placed backward—are these anomalies in your training data? Probably not.

This is a widespread problem in robot learning: existing datasets almost exclusively contain successful execution records. Failed, anomalous, and partially completed executions—data essential for training robust systems—are severely missing.

COFFAIL: Recording Both Success and Failure

The COFFAIL (Coffee Failure) dataset fills this gap. It records the Jessie robot executing seven different skills in coffee-preparation scenarios, including both successful and anomalous executions.

The seven skills cover the complete coffee-making workflow: picking up cups, operating the coffee machine, adding milk, stirring, delivering cups, and more.

What makes the dataset unique:

Anomaly diversity: Not just "it failed," but recordings of many different failure types—grasping slips, placement offsets, wrong action ordering, too much or too little force, etc.
Real physical interaction: Not simulation data, but real recordings of a real robot in a real environment.
Complete episodes: Each record contains the full execution from start to finish, including both successful and failed moments.

Why Failure Data Matters

From a machine learning perspective, training only on successful data has serious problems:

Distribution shift: In real deployment, anomalies are unavoidable. If a model has never seen failure, it doesn't know how to recover—or even that it has failed.

Fragility: Models trained only under "ideal conditions" are extremely sensitive to noise, disturbances, and small deviations. A successful coffee-making motion can fail entirely if the cup is 1 cm off.

Safety risks: A robot operating in the real world that doesn't know what "anomalous" looks like cannot trigger safety mechanisms. Recognizing anomalies is the first step of any safety system.

COFFAIL enables researchers to:

Train anomaly detection models so robots know "when something is wrong"
Train recovery policies so robots can "rescue themselves" from failure
Evaluate model robustness rather than just success rates

My Take

Although the paper is short (the authors themselves call it a "short paper"), it makes an important point: in robot learning, we should not focus only on success—we should systematically collect and analyze failure.

It reminds me of aviation safety. An aircraft's black box records all flight data, including accidents and near misses. Much of aviation's safety progress has come from deep analysis of "failure cases," not from repeatedly practicing successful flights.

Robot learning should work the same way. COFFAIL, though limited to a coffee-preparation scenario, represents a methodology: systematically recording, classifying, and learning from failure is a necessary path toward robust robotic systems.

And honestly, a dataset called COFFAIL is inherently fun—who wouldn't want to see what a robot failing at coffee-making looks like?

---

Paper: arxiv.org/abs/2604.18126 Robot: Jessie Robot