Evaluated Challenges

The agents developed in Duckietown Learning Experiences can be evaluated against a set of benchmarks defined as a Duckietown Challenge.

Learners may evaluate their agent locally via

dts code evaluate

or on the Duckietown Challenges Server via

dts code submit

In each of these cases, a report is created against these benchmark metrics along with visualizations of the agent’s behavior in simulation. Server submissions are appended to a running leaderboard, allowing learners to compare their solutions with previous work.

You can explore previous challenge definitions and the related student submissions on the Duckietown Challenges Server.

Fig. 27 Each challenge and linked learner submissions can be found on the Duckietown Challenges Server.

Development Guidelines

Communicating benchmarks:

It is recommended that the evaluation metrics for a challenge are clearly defined in the Grading section of the Learning Experience README.md file to give learners clear performance goals for their agent.

Encouraging frequent commits:

Students should also be encouraged to bookmark their development at each submission attempt with a git commit referencing the submission number. This allows them to easily track and revert to prior attempts.

Tutorial: Creating and Linking a Duckietown Challenge

Learning Experiences can be associated with a predefined challenge by updating the fields in the file recipe/submission.yaml.

Attention

The ability to create custom challenges for a Learning Experience has not yet been released. Please check back later for future updates.