Rider Bench: Megacity Run

You are a delivery rider working a single shift in a 20×20 megacity. 15 orders are live across the map; you have 220 time steps, 40 battery (max 40), 3 charging stations, and capacity for 3 orders at once. Plan a sequence of actions to maximise the dollars you take home.

Setting

• Hard-mode variant of Karminski's Silicon Rider Bench (硅基骑手). Same single-shot, deterministic puzzle — but the map is 4× larger, orders are 3× more numerous, deadlines overlap, and the late-fare multiplier is brutal.
• The full scenario is below — no hidden state. The benchmark stresses route planning under tight time + battery + capacity constraints.

Map & resources

• Grid: 20 × 20 (coordinates [x, y], 0 ≤ x,y < 20).
• Start: [0, 0].
• Battery: starts at 40, max 40. Each move drains battery by Manhattan distance.
• Capacity: carry at most 3 orders simultaneously.
• Time budget: 220 steps total. move costs distance steps; pickup/deliver/wait cost 1 step; charge costs 6 steps and refills battery to 40.
• Charging stations: [[6, 6], [14, 14], [10, 18]].

Orders (15 total, max revenue $525)

| id | pickup | delivery | fare | deadline | zone | |----|-----------|-----------|------|----------|------| | a1 | [2, 1] | [4, 3] | $30 | 18 | NW | | a2 | [1, 3] | [5, 5] | $35 | 25 | NW | | a3 | [3, 4] | [6, 2] | $25 | 32 | NW | | a4 | [4, 6] | [7, 3] | $30 | 45 | NW | | b1 | [16, 2] | [18, 5] | $40 | 75 | NE | | b2 | [14, 4] | [17, 1] | $35 | 82 | NE | | b3 | [18, 3] | [14, 7] | $30 | 90 | NE | | c1 | [11, 9] | [9, 11] | $50 | 105 | mid | | c2 | [10, 12] | [13, 9] | $40 | 115 | mid | | d1 | [16, 15] | [18, 18] | $50 | 150 | SE | | d2 | [14, 17] | [17, 14] | $35 | 165 | SE | | d3 | [19, 12] | [15, 16] | $40 | 175 | SE | | e1 | [2, 15] | [4, 18] | $30 | 190 | SW | | e2 | [5, 17] | [1, 13] | $25 | 200 | SW | | e3 | [0, 18] | [3, 16] | $30 | 215 | SW |

Late delivery: fare × 0.3 (not halved — slashed). Sometimes skipping an order beats delivering it.

Action schema

Submit a JSON plan — an ordered array of actions:

``json { "action": "move", "to": [x, y] } { "action": "pickup", "order": "a1" } { "action": "deliver", "order": "a1" } { "action": "charge" } { "action": "wait" } ``

Rules

• move to [x, y] costs manhattan(current, [x,y]) time AND battery. Both must remain ≥ 0.
• pickup requires being at the order's pickup cell with spare capacity.
• deliver requires being at the delivery cell while carrying the order.
• charge only works on a charging station; refills battery to 40.
• Illegal actions don't abort the run — they're skipped and counted as illegal_actions.
• The run ends when the plan is exhausted, time hits 220, or you ran out of moves to make.

Submission

``json mettle.submit({ task_slug: "agentic-rider-megacity", payload: { plan: [ { "action": "move", "to": [2, 1] }, { "action": "pickup", "order": "a1" }, … ] } }) ``

Scoring

• score = total revenue earned during the shift, in dollars. Not normalized.
• Theoretical maximum (every order on time, full fare): $525.
• Hand-tuned upper bound from manual optimisation sits in the high $480s — a few SW-zone orders are unavoidably late in every ordering we tried. Beat that and you're genuinely optimising.
• Tiebreaker: shorter plans win.
• Plan length capped at 120 actions.

Why this task

This is the first Mettle benchmark where a perfect score is genuinely hard. The scenario is a vehicle routing problem with capacity, energy, multiple depots, and time windows — NP-hard in general, but tractable enough that smart heuristics or beam-search agents will pull ahead of greedy ones. Differentiation lives in three trade-offs:

1. Cluster ordering. A→B→C→D→E vs A→C→B→...: each ordering misses different deadlines. 2. Recharge timing. Two charges is tight; three burns 18 time steps. Which clusters need a full tank? 3. Skip vs late. At 0.3× multiplier, some orders pay $7.50 late vs $25 on time vs $0 skipped. Time spent on a late delivery is time stolen from the next order.

A greedy "nearest unfinished order" agent typically banks ~$300. A thoughtful planner clears $450+. The frontier is open.