| --- |
| license: cc-by-nc-4.0 |
| pretty_name: pack_route tasks |
| tags: |
| - reinforcement-learning |
| - rl-environment |
| - agents |
| - tool-use |
| - operations-research |
| - vehicle-routing |
| - bin-packing |
| size_categories: |
| - n<1K |
| configs: |
| - config_name: v1 |
| default: true |
| data_files: |
| - split: train |
| path: data/v1/train.jsonl |
| - split: test |
| path: data/v1/test.jsonl |
| - config_name: v0 |
| data_files: |
| - split: test |
| path: data/v0/test.jsonl |
| --- |
| |
| # pack_route tasks |
| |
| 100 tasks for the [pack_route](https://github.com/savarbhasin/pack-route) RL environment (Verifiers v1). In each task |
| an agent is the dispatcher at a parcel station. It decides which parcels ride in which van **and** the order each van |
| visits its stops, to minimise total driving time. The stops, parcels and travel times come from real 2018 Amazon Last |
| Mile routes. The van fleet is synthetic. |
| |
| Packing the vans first and routing second is a trap: it loses to the joint pack+route reference on every task. So an |
| agent only reaches 1.0 by planning packing and routing together. |
| |
| | Config | Split | Tasks | Stops / vans | Reference | |
| | --- | --- | --- | --- | --- | |
| | `v1` (default) | `train` | 80: 20 each from Los Angeles, Seattle, Chicago, Boston | 20/3, 25/3, 30/4, 40/4, mixed | best found | |
| | `v1` | `test` | 10 from Austin, a city `train` never sees | same mix | best found; **plans removed** | |
| | `v0` | `test` | 10 (the first release) | 14/3 | proven optimum | |
| |
| ```python |
| from datasets import load_dataset |
| train = load_dataset("dev-front/pack-route-tasks", "v1", split="train") |
| ``` |
| |
| ## How a task is made |
| |
| 1. Take one real route: one van's day from one station (2018-07-20 to 2018-08-24 in v1). |
| 2. Pick 20–40 of its stops, spread across the route, with their real parcels (23–155 per task) and the real driving |
| times between them and the station. |
| 3. Add a synthetic fleet of 3 or 4 vans, sized so the parcels fill 85–95% of it. |
| 4. Solve it with CP-SAT, jointly packing and routing. At this size a laptop rarely proves the optimum, so v1 uses the |
| best plan found: the search stops after 30 s without improvement (at most 180 s), and the cheapest of its plan and |
| the baselines becomes the reference. The solver's lower bound is stored too. |
| 5. Keep the task only if the best "pack first, then route" plan grades ≤ 0.9 against the reference. |
| |
| Every task comes from a different route. Every reference plan was checked with the environment's validator at exactly |
| the solver's cost. |
| |
| ## Fields |
| |
| | Field | Meaning | |
| | --- | --- | |
| | `id` | `<route>-spread-n<stops>-k<vans>-t<fill>-s<seed>` | |
| | `instance.stops` | Stop ids `S01`… The depot is row/column 0 of `travel_s`, and stops follow in this order | |
| | `instance.parcels` | `id`, `stop`, `volume_dl` (volume in decilitres, rounded from real dimensions) | |
| | `instance.vans`, `instance.capacity_dl` | Fleet size and per-van capacity. **Both are synthetic** | |
| | `instance.travel_s` | Real historical driving times in seconds; asymmetric | |
| | `reference.optimum_s` | The reference total driving time the grade is measured against | |
| | `reference.optimality` | `proven` (v0: the proven optimum) or `best_found` (v1: the best plan known) | |
| | `reference.lower_bound_s` | No plan is cheaper (CP-SAT's bound; median 13% below `optimum_s` in v1) | |
| | `reference.floor_s` | Depot legs no plan can avoid; used to normalise the gap | |
| | `reference.ffd_s`, `reference.stop_aware_best_s`, `reference.stop_aware_mean_s` | Pack-then-route baseline costs | |
| | `reference.plans` | The reference, FFD and best stop-aware plans in the environment's plan format; `null` in `v1` test | |
| | `meta` (v1) | Route id, city, station, date, stops, vans, fill, build seconds, and the stop-aware baseline's grade | |
|
|
| **`reference` is the answer key.** The environment never shows it to the agent; keep it out of prompts. The `v1` test |
| split keeps the costs needed for grading but not the plans. |
|
|
| ## Grade (PortSim reward v3) |
|
|
| | Plan | Score | |
| | --- | --- | |
| | Valid | `0.2 + 0.8·e^(−g/0.5)`, with `g = max(0, cost − optimum_s) / (optimum_s − floor_s + 100)` | |
| | Invalid | `0.2 ×` share of cleanly delivered parcels | |
| | None submitted | 0 | |
|
|
| Matching or beating the reference scores 1.0. In v1 a model can beat it, since it is not proven optimal. |
|
|
| Baselines on v1: the best stop-aware plan (keep each stop in one van where possible, then route each van optimally) |
| scores 0.29–0.79 (mean 0.58, 11% more driving). First-fit-decreasing scores 0.25–0.70. On v0 FFD scores 0.32–0.67. |
|
|
| ## Not modelled yet |
|
|
| Time windows, service time, driver shifts and mid-episode news. |
|
|
| ## Licence and attribution |
|
|
| Derived from the [2021 Amazon Last Mile Routing Research Challenge dataset](https://registry.opendata.aws/amazon-last-mile-challenges), |
| CC BY-NC 4.0, and released under the same licence (**non-commercial**). Cite: Merchán et al., "2021 Amazon Last Mile |
| Routing Research Challenge: Data Set", *Transportation Science* (2022), |
| [doi:10.1287/trsc.2022.1173](https://doi.org/10.1287/trsc.2022.1173). The reward formula and the "real records + solver |
| optimum" pattern come from [PortSimEnv](https://github.com/adithya-s-k/FineEnvs/tree/main/07-simulation-environments) |
| (FineEnvs). |
|
|
