Fleet Recovery

Fleet Recovery

Two autonomous couriers share a warehouse floor. They share something else as
well: the same compact neural policy. Can one model coordinate both robots,
deliver every package, survive actuator failures, and still generalize when the
warehouse geometry changes?

The mission

Each hidden episode is a 10×10 grid with two robots, four pickup/delivery pairs,
two chargers, and eighteen blocked cells. A robot may move in the four cardinal
directions, pick up, drop off, or wait. Robots cannot occupy the same cell or
swap cells in one step. Move attempts consume battery; a robot stranded at zero
battery away from a charger is lost for the episode.

Actuators sometimes stall. A movement command then consumes energy without
changing position. The observation exposes the stalled state; wait performs a
recovery. Waiting on a charger also restores energy. The policy receives the
other robot's state, package state, compact route features, and a conservative
right-of-way hint from the fixed navigation stack.

Your model

Train or construct one shared ReLU MLP:

172 inputs → 64 ReLU → 32 ReLU → 7 logits

At every step the evaluator runs the same weights independently for robot 0 and
robot 1, masks illegal actions, then executes both decisions simultaneously.
Only fixed float32 weights are accepted; participant code is never executed.

The public archive contains expert demonstrations, familiar validation layouts,
a NumPy baseline, an exporter, and a validator. The compact features support
several approaches: supervised imitation, data reweighting, synthetic
augmentation, reinforcement learning on a reimplemented simulator, policy
distillation, or direct weight synthesis.

Distribution shift

• Partial score: familiar scattered obstacles and ordinary short stalls.
• Complete score: unseen corridor/room layouts, exchanged starting roles,
  tighter battery budgets, and more persistent stalls.

Each split has 24 private scenarios and every rollout lasts exactly 160 steps.

Score

Deliveries determine most of the score. Earlier deliveries and remaining energy
add smaller bonuses. Collisions, battery failures, and repeated 16-step
no-progress periods are penalized. The exact formula and feature indices are in
the dataset README, including the public continuous band calibration used to
make improvements in the functional range visible. Scores lie in [0, 100];
higher is better.

Suggested milestones: 25+ functional, 50+ competitive, 75+ excellent.

Submission format

Upload one ZIP containing exactly manifest.json and policy.npz. The six NPZ
arrays and their required shapes are documented in the participant kit. Arrays
must be finite float32 values bounded by 25 in absolute value. Maximum archive
size: 256 KB after platform extraction of the outer ZIP.