TTP-D solver weights
Paper: Drive, Pack, Fly: The Travelling Thief Problem with Drone
Trained policy checkpoints for the study Fly, Pack, Drive: the Travelling Thief Problem with Drone.
A capacitated truck and a single-package drone operate from a common depot on a collection route. The truck's velocity decreases affinely with its accumulated load, so an early pickup penalises every subsequent arc. The drone launches from a node at which the truck is present, retrieves one item at an outlying customer, and rejoins the truck at a later rendezvous node; whichever vehicle arrives first waits, while the rental clock continues to run. The objective is to maximise collected profit net of a rental cost proportional to the makespan.
These checkpoints are the learned construction policies: an encoder–decoder model that embeds the instance once and emits a composite action at each node the truck reaches, trained offline with Proximal Policy Optimisation under a POMO group baseline.
Code: https://github.com/corbit-lab/ttpd. Benchmark instances, result tables,
and behaviour-cloning datasets are in the companion dataset repository
Murjani/ttpd-benchmarks.
Layout
<variant>/<model>/<family>/n<N>/best.pt
Variant is a280 (fixed drone endurance) or ttd300 (endurance-conditioned).
Model is gat (attention encoder), mlp (the encoder ablation), or lisa
(behaviour cloning). Family denotes the distribution the policy was trained
on, which determines its applicability:
| Family | Training distribution | Intended use |
|---|---|---|
benchmark-tuned/ | The five benchmark instances of that size | The reported headline results |
sampled/ | Randomly sampled a280 subsets | Generalisation; warm start for the above |
specialists/ | Early small-N runs | The N ≤ 20 rows, using dedicated small models |
The benchmark-tuned policies warm-start from sampled/n20 of the same model
family, as training from scratch does not converge within the budget at the
larger sizes. Within a run, best.pt is the deliverable, selected by held-out
beam evaluation; last.pt supports resumption; and best_milp.pt is selected
by MILP gap rather than evaluation return.
Behaviour-cloned policies are stored at
<variant>/lisa/behaviour_cloning/n<N>.pt, with
a280/lisa/behaviour_cloning_initial/ and a280/lisa/gat-n50/ retained for
provenance.
Deliverable checkpoint per model and size
| Model and size | Checkpoint |
|---|---|
| a280 GAT, N = 5, 10 | a280/gat/specialists/n5.pt, n10.pt |
| a280 GAT, N = 15, 20 | a280/gat/sampled/n<N>/best.pt |
| a280 GAT, N = 30–100 | a280/gat/benchmark-tuned/n<N>/best.pt |
| a280 MLP, N = 20 | a280/mlp/sampled/n20/best.pt |
| a280 MLP, remaining sizes | a280/mlp/benchmark-tuned/n<N>/best.pt |
| ttd300 GAT | ttd300/gat/benchmark-tuned/n<N>/best.pt |
| LISA | <variant>/lisa/behaviour_cloning/n<N>.pt |
The benchmark-tuned family also carries the smaller sizes so that a
directory-based checkpoint argument resolves uniformly.
Loading
Checkpoints are torch.save payloads keyed under "policy":
import torch
from huggingface_hub import hf_hub_download
path = hf_hub_download("Murjani/ttpd-weights", "a280/gat/benchmark-tuned/n30/best.pt")
payload = torch.load(path, map_location="cpu", weights_only=False)
policy.load_state_dict(payload["policy"])
The repository code resolves these paths automatically through
ttpd.hub.ensure_local, which returns an existing local file unchanged and
downloads otherwise, so a machine that already holds the tree performs no
network access.
manifest.json records, for every file, the path from which it was published
together with its SHA-256 digest.
Citation
Please cite the accompanying paper. Reported results, including the comparison against the exact solver and the metaheuristics, are given there.
@inproceedings{murjani2026ttpd,
title={Drive, Pack, Fly: The Travelling Thief Problem with Drone},
author={Murjani, Kabir and Sobhanan, Abhay},
year={2026},
eprint={2608.16435},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2608.16435},
}