Checklist AICC 2026 Round 8 · Task 2
Pixel Quest
From 64 × 64 first-person raycast frames, predict the robot's heading change between two frames and train an agent that navigates to a target in a Gymnasium environment.
The task
A robot in a 20 × 20 grid world sees only a 64 × 64 RGB first-person view rendered by raycasting, with a compass in the bottom-left corner pointing towards a golden target. Grass and sand are walkable; water acts as a wall. The robot can move forward, turn left, turn right or move backward.
Subtask 1 (visual odometry, 40 points): 400 training and 100 test pairs of before/after frames are given, labelled with dx, dy (each −3 to +3) and dheading (0 no turn, 1 right, 2 turn around, 3 left). Only the heading prediction is scored.
Subtask 2 (pixel navigation, 60 points): there is no pre-made dataset; the agent must be trained in the PixelNav-v0 Gymnasium environment. Episodes start in a random safe cell with the target at least 6 cells away and last at most 100 steps. Rewards are −1 per step, −5 for hitting water or the world edge, +0.5 per cell moved closer to the target and +100 for reaching it. Contestants submit an action list for each of 100 hidden test episodes, which is replayed in the environment.
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
- Starter kit with the pixelnav environment package and
baseline.ipynb(which must be used); training-data zip with the odometry pairs ({pair_id}_before.png,{pair_id}_after.png) andlabels.csv(pair_id, dx, dy, dheading). - You submit
submission.csvwith columns subtaskID, datapointID, answer:odom_{pair_id} rows with 'dx;dy;dheading' andnav_{episode_id} rows with a comma-separated list of action numbers.- Scoring
- Subtask 1: heading accuracy; 0 points below 30%, linear from 0 to 40 points between 30% and 90% (points = 40 × (accuracy − 0.30) / (0.90 − 0.30)), 40 points at 90% or more. Subtask 2: mean total reward over the 100 test episodes; 0 points at −110 or below, 20 × (reward + 110) / 110 between −110 and 0, 20 + 40 × reward / 50 between 0 and 50, 60 points at 50 or above. Maximum 100.
- Rules
- The provided
baseline.ipynbmust be used (it contains the dataset and submission boilerplate). - Each submission must include the source code as a .py file (maximum 35 KB) and the predictions CSV (maximum 50 MB); organiser-provided seed-setting code must be used.
- Round rules (honour system): no LLMs or AI assistants for code or ideas; no internet beyond official library documentation; no communication or sharing of approaches during the contest; individual participation only.
- The provided
- Format
- AICC Round 8 (special edition following the CEOAI syllabus), online on Nitro AI Judge; open from 12 Jun 2026 17:00 to 14 Jun 2026 17:00 (time zone not stated on the platform); contestants are asked to work within a single self-chosen 6-hour session.