Checklist ROAI 2026 Selection Camp – CPU Practical Round · Task 3
Smart Warehouse
Implement TD(0), SARSA and Q-learning updates and train a tabular agent that picks up and delivers a package in hidden SmartWarehouse-v0 grid scenarios.
The task
SmartWarehouse-v0 is a rectangular grid in which a robot must reach a package, pick it up and deliver it before its battery is exhausted or the step limit is reached. Walls block movement, dangerous cells add a penalty, some scenarios include slip probability, and a recharge cell restores the battery. Observations are integers encoding (row, column, battery, package status); there are seven discrete actions.
Rewards are +50 for delivery, +8 for pickup, −1 per action, −3 for invalid actions, −8 for entering a dangerous cell and −20 when the battery empties. The contestant submits solution.py and solution.pkl defining td_update, q_learning_update, sarsa_update, train_agent and select_action; the evaluator unit-tests the update rules, trains the agent on hidden scenarios and evaluates the greedy policy.
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
- Public warehouse scenarios for local testing and a restricted environment interface (reset, step, action/observation space sizes,
max_steps, decode). - You submit
solution.pyandsolution.pklimplementing the five required functions.- Scoring
- 100 points: TD(0) update tests 10, SARSA update tests 12, Q-learning update tests 13, terminal-state handling 10, required interface 5, hidden-scenario normalised return 25 (normalised between a random baseline and a teacher solution), hidden-scenario success rate 25.
- Rules
- CPU only
- TRAIN_EPISODES = 8000, EVAL_EPISODES = 40
- Training timeout 12.0 s per hidden scenario; 0.20 s per
select_actioncall - Private environment fields (env.scenario, env._walls, env._danger) must not be used
- On-site at the National College of Informatics "Tudor Vianu", Bucharest
- Internet access and packages restricted to the allowed list in the round rules
- Final score computed on hidden test data; only selected submissions count
- Format
- ROAI 2026 selection camp, CPU practical round, 28 May 2026.