Robot dataset datasheet template
A datasheet is what a person reads before deciding to train on your data. The structure follows the questions in "Datasheets for Datasets" (Gebru et al., Communications of the ACM, 2021), adapted for robot data: embodiment, streams, synchronization and the gap between how the data was collected and where a policy will run.
Who fills it: The dataset owner, at each release.
When to use it
- Releasing a dataset publicly or to another team.
- Freezing a dataset version that models will be trained and evaluated on.
- Handing a dataset to a customer or partner who wasn't involved in collecting it.
When something else fits better
- Collection hasn't started: write the data requirements spec. The datasheet describes what you got, not what you planned.
Preview
1.Motivation
Why the dataset was created and what gap it fills.
Who created it and who funded it.
2.Composition
One attempt at which task, from start to end state.
Episodes, hours, successes and failures, per task, environment and collection method.
Robot(s) and end effector(s), or human hands with which wearable sensors.
Each stream with modality, device, rate, units and frame. Link to the annotation schema.
Which labels exist (success, segments, contact events) and how they were produced.
Recommended train, validation and held-out splits, and what is held out (environments, objects, operators).
Whether episodes contain faces, voices, biosignals or other personal data, and how they were handled.
3.Collection process
Teleoperation, kinesthetic, human demonstration with wearables, scripted or autonomous, with shares.
Link the collection protocol version(s) used.
Clock source and the measured sync offsets across sessions (range).
How many operators, how they were selected and trained, and how consent was obtained.
When and where data was collected.
4.Preprocessing and cleaning
Filtering, resampling, cropping, anonymization, episodes removed and why.
5.Uses
The tasks and model types the dataset was built for.
Uses the data is unsuitable for or not permitted for.
Example: Not suitable for tasks with human contact; not licensed for identifying operators.
Policies trained on it and their evaluation results, linked to evaluation plans.
6.Limits and known issues
Conditions, objects or behaviours the data doesn't cover.
Known labeling errors, sync problems or sensor faults, and which episodes they affect.
Differences between the embodiment in the data and target robots.
7.Distribution and maintenance
License, access conditions and where to download.
Current version and what changed from the previous one.
How to report problems and where corrections are published.
Copy or download
The whole template with prompts, examples and the reason for each field.
# Robot dataset datasheet
A datasheet is what a person reads before deciding to train on your data. The structure follows the questions in "Datasheets for Datasets" (Gebru et al., Communications of the ACM, 2021), adapted for robot data: embodiment, streams, synchronization and the gap between how the data was collected and where a policy will run.
Who fills it: The dataset owner, at each release.
How to use this file: write your answer after each "Answer:". The "Why" notes explain each field for reviewers; keep them or delete them once the document is agreed.
## 1. Motivation
### 1.1 Purpose
Why the dataset was created and what gap it fills.
Answer:
> Why: Tells readers the problems it was designed for, which are the ones it is most likely to work on.
### 1.2 Creators and funding
Who created it and who funded it.
Answer:
> Why: Lets readers judge possible conflicts of interest and know whom to contact.
## 2. Composition
### 2.1 What an episode is
One attempt at which task, from start to end state.
Answer:
> Why: Readers need the unit before they can interpret any count.
### 2.2 Counts
Episodes, hours, successes and failures, per task, environment and collection method.
Answer:
> Why: Totals hide imbalance. A breakdown shows what the data covers well and where it is thin.
### 2.3 Embodiment
Robot(s) and end effector(s), or human hands with which wearable sensors.
Answer:
> Why: Determines how directly the data can supervise a given robot.
### 2.4 Streams
Each stream with modality, device, rate, units and frame. Link to the annotation schema.
Answer:
> Why: Lets readers check the data contains the signals their policy needs.
### 2.5 Labels
Which labels exist (success, segments, contact events) and how they were produced.
Answer:
> Why: Automatic and human labels have different error profiles. Readers need to know which they are using.
### 2.6 Splits
Recommended train, validation and held-out splits, and what is held out (environments, objects, operators).
Answer:
> Why: Shared splits make results comparable across users of the dataset.
### 2.7 Personal data
Whether episodes contain faces, voices, biosignals or other personal data, and how they were handled.
Answer:
> Why: Users need to know their obligations before they download.
## 3. Collection process
### 3.1 Method
Teleoperation, kinesthetic, human demonstration with wearables, scripted or autonomous, with shares.
Answer:
> Why: The method shapes the behaviour in the data: operator style, speed, recovery strategies.
### 3.2 Protocol
Link the collection protocol version(s) used.
Answer:
> Why: Lets readers see exactly how sessions were run.
### 3.3 Synchronization
Clock source and the measured sync offsets across sessions (range).
Answer:
> Why: Measured offsets tell users how precisely contact events line up with video and actions.
### 3.4 Operators and consent
How many operators, how they were selected and trained, and how consent was obtained.
Answer:
> Why: Operator variety affects behaviour diversity; consent terms limit allowed uses.
### 3.5 Time frame and sites
When and where data was collected.
Answer:
> Why: Hardware revisions and environments change over time; dates help explain shifts.
## 4. Preprocessing and cleaning
### 4.1 What was done
Filtering, resampling, cropping, anonymization, episodes removed and why.
Answer:
> Why: Every transformation is a decision a user may want to undo. Say whether raw data is available.
## 5. Uses
### 5.1 Intended uses
The tasks and model types the dataset was built for.
Answer:
> Why: Sets expectations: where the creators expect it to work.
### 5.2 Uses to avoid
Uses the data is unsuitable for or not permitted for.
Example: Not suitable for tasks with human contact; not licensed for identifying operators.
Answer:
> Why: Stated limits prevent misuse and save users from wasted training runs.
### 5.3 Results so far
Policies trained on it and their evaluation results, linked to evaluation plans.
Answer:
> Why: Real results, with their evaluation setup, are the best evidence of what the data supports.
## 6. Limits and known issues
### 6.1 Coverage gaps
Conditions, objects or behaviours the data doesn't cover.
Answer:
> Why: Every dataset has gaps. Listing them is more useful than an impression of completeness.
### 6.2 Known errors
Known labeling errors, sync problems or sensor faults, and which episodes they affect.
Answer:
> Why: Lets users filter affected episodes rather than discover the problem in their results.
### 6.3 Embodiment gap
Differences between the embodiment in the data and target robots.
Answer:
> Why: The most common reason robot data underperforms on a new platform.
## 7. Distribution and maintenance
### 7.1 License and access
License, access conditions and where to download.
Answer:
> Why: Users must know what they may do with it before they invest in it.
### 7.2 Versioning and changelog
Current version and what changed from the previous one.
Answer:
> Why: Results must cite a version; a changelog explains why numbers moved.
### 7.3 Contact and errata
How to report problems and where corrections are published.
Answer:
> Why: A maintained dataset improves; an unmaintained one accumulates known issues nobody hears about.