Before you begin
Starting knowledge
- Basic Python, arrays and image processing.
- Understand the difference between a sensor observation and ground truth.
Equipment & environment
Start with licensed existing data and a controlled scene. Capture identifiable people only under an appropriate consent and governance process.
No fixed completion time: use the evidence gates below to decide when to advance. Resource links open the botstest directory entry; check its publisher documentation for the exact version and configuration.
Beginner
Define observations and labels
Connect a perception question to measurable annotations.
Learn in this order
- Detection, pose and geometry task definitions
- Camera intrinsics, frame conventions and timestamps
- Label provenance, ambiguity and permissions
Put it into practice
- Inspect a small image or point-cloud sample and its labels.
- Write annotation rules for difficult and unobservable cases.
A data dictionary, annotation guide and sample audit.
Ready for the next step when…
Resources for this stage
Inspect point clouds and geometry-processing results before using them in a perception pipeline.
Study image processing and camera-related tooling through its upstream project documentation.
Inspect hand-object observations, annotations and evaluation assumptions for a focused perception study.
Intermediate
Construct leakage-resistant evaluation
Measure model behavior on a meaningfully held-out distribution.
Learn in this order
- Scene, object and session splits
- Calibration quality and sensor alignment
- Precision/recall or task-appropriate error definitions
Put it into practice
- Split by capture session or object where the task requires it.
- Compare a simple baseline on normal, occluded and shifted observations.
A versioned dataset split and baseline evaluation notebook.
Ready for the next step when…
Resources for this stage
Inspect a manipulation dataset and its collection conditions before designing training and evaluation splits.
Study a manipulation data-collection workflow and its embodiment and environment assumptions.
Explore fiducial detection and pose observations with known marker geometry and camera calibration.
Professional
Release a dataset with an honest scope
Make the data reusable and its limitations visible.
Learn in this order
- Dataset documentation and usage rights
- Sensor and scene distribution reporting
- Model confidence, drift and downstream consequences
Put it into practice
- Prepare a data card with collection, processing and split details.
- Have a second researcher reproduce evaluation from the released metadata.
A dataset card, reproducible evaluator and limitations report.
Ready for the next step when…
Resources for this stage
Examine multi-embodiment demonstration data, metadata, licenses and action conventions.
Explore object assets with provenance and license checks before including them in simulation.
Visualize spatial observations and compare them with model and trajectory outputs.
Common pitfalls
- Randomly splitting adjacent video frames and calling the result generalization.
- Using rendered or human-video data as if it were robot-action ground truth.
Where to go next
Extend to manipulation or industrial inspection with the same data and evaluation contract.
Explore other research paths →