Checklist IOAI 2026 At-Home Round · Task 2
Robot Delivery Academy
Use behavioural cloning on expert demonstrations to train a model that drives a delivery robot to pick up and deliver packages.
The task
A small delivery robot works on an 8×8 city map with blocked cells and six depots labelled A to F. In each episode it must reach the depot holding the package, pick it up, carry it to the destination depot and drop it off, and each map is slightly different.
Instead of hand-writing a route planner, you learn the behaviour from a deliberately small set of expert demonstrations, as supervised observation-to-action examples. The trained model then drives complete episodes step by step on validation and test scenarios it has not seen. One wrong move can push the robot into situations that were rare in the demonstrations, so good per-action accuracy does not guarantee successful deliveries.
Each observation combines a 6×8×8 grid of feature planes, a vector of 13 normalised values and a mask of currently valid actions. The six actions are south, north, east, west, pickup and dropoff. The notebook includes a simulator for replaying and visualising episodes.
Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.
At a glance
- You get
train_demos.pkl(expert trajectories with observations and actions), plusvalid_scenarios.pklandtest_scenarios.pkl(scenarios without action labels), downloaded from a Google Drive folder. The notebook's saved outputs show 400 training demonstrations (5,327 state-action samples), 200 validation scenarios and 1,600 test scenarios.- You submit
- A notebook that produces
predictions.zipcontainingpredictions.jsonl, with one line per test scenario: {"layout_id", "episode_seed", "actions": [...]}, using action IDs 0–5. The evaluator replays the actions from each scenario's start state. - Scoring
- Success rate: the number of episodes where the package is delivered to the destination within the step limit, divided by the total number of episodes.
- Rules
- Train on the provided demonstrations; do not use expert action labels for validation or test scenarios
- Do not generate extra expert trajectories with search, planning or another expert model
- The final prediction process must be deterministic
- The submitted notebook must generate
predictions.zipfrom scratch - Rule-based or hard-coded solutions may be reviewed by the Scientific Committee
- Format
- At-Home Round (educational; the 2026 Contest Rules say the three problems were given to teams about one month before IOAI 2026 and do not affect final results)