Discord

Checklist IOAI 2026 Individual Contest · Task 2

Robot Chasing

Predict the next action of each of six grid-world robots from a snapshot of its room and its English mission.

  • Next-action classification

The task

Six robots each work in a small room: a 6×6 floor ringed by walls, stored as an 8×8 grid. Each robot gets an English mission of one of three kinds (go to an object, pick one up, or place one object beside another), and a mission can be worded in several ways. Every robot follows its own consistent habits and does not always take the shortest route.

A sample is a single snapshot taken at some point during a mission. It gives the robot ID, a two-channel grid of object and colour codes, the direction the robot faces, the mission text and anything it is carrying. Snapshots are shuffled and never linked into episodes. Test missions may combine familiar words, colours and objects in new ways, but every word, phrase pattern, colour, object type and mission type also appears in training.

For each snapshot you predict the robot's next action out of six: move up, down, left or right (a move also turns the robot, even if blocked), pick up, or drop.

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
observations.json lists of snapshots (robot_id, 8×8×2 image of object and colour indices, direction, mission, carrying). dataset/train has 60,000 rows with labels.json (10,000 per robot and 20,000 per mission type) and dataset/test_public has 3,600 rows (600 per robot). At grading time test_public is replaced by a hidden 3,600-row set without labels: test_leaderboard_a for the public leaderboard, test_leaderboard_b for the final ranking.
You submit
predictions.json: a JSON list with one integer action (0–5) per row of dataset/test_public/observations.json, in the same order. A missing or invalid file, a wrong number of predictions, a non-integer or an out-of-range action is rejected without a score. Submit exactly one file, solution.ipynb.
Scoring
Mean per-robot accuracy on a 0–100 scale: accuracy is computed separately for each of the six robots, then averaged so every robot has equal weight.
Rules
  • Time limit 5 minutes
  • One GPU (≈16 GB VRAM), no internet, 5 GB storage
  • solution.ipynb ≤ 1 MB
  • Labels may be read only from dataset/train/ (a notebook that unconditionally reads test labels will fail)
Format
Individual Contest, Day 1 (the 2026 Contest Rules give six hours for the day's three tasks); the same task was set in the GAITE Contest with a hint

Details

Year
2026, Astana, Kazakhstan
Rounds
Individual Contest · Task 2 · Day 1GAITE Contest · Task 2 · Day 1
Language
English
License
CC BY 4.0, as stated by the source