Learned policies for classical planning act without search, but each is tied to the domain it was trained on: network modules per action schema, parameters per predicate, or vocabulary-sized encodings mean a trained policy cannot even be executed on a new domain's predicates and action schemas. We present GADAR (Generalizing Across Domains via Action Ranking), a policy that takes the domain description as input, π(s, G, D), and constructs grounded actions in domains never seen in training, without search. Symbols are described by the roles they play in the domain rather than by their names, in three layers that each extend the previous one: (i) a lifted domain layer embeds D in every state graph through structural signatures; (ii) a binding layer ties each ground action's applicability to the lifted preconditions it instantiates; and (iii) a chain layer states which schema produces what another needs and how far each schema stands from the goal. The same network therefore runs on every domain, seen or unseen. We evaluate zero-shot generalization leave-one-domain-out over eight benchmark domains. A single GADAR policy trained on seven domains solves 58.3% of their test instances, one point below the 59.3% of dedicated single-domain GADAR models, so one policy serves many domains at no cost in coverage; a multi-task control that keeps symbol identity in its weights, while strong on some of the domains it trains on, transfers to none of the held-out ones. Zero-shot, GADAR executes on every held-out domain and exceeds random choice under the same executor on the large test instances of all eight: it solves 22.6% of them on average, against 3.3% for random choice and 0.6% for the same representation without the chain layer, and reaches 68% on held-out Visitall. Zero-shot coverage nonetheless remains well below in-domain coverage on most domains, and we analyze where it is weakest.
Figures & tables
Figure 1: The GADAR representation for a Blocksworld task. The ground instance layer (bottom) is the action-centric graph of the grounded policy we build on ( Mangannavar et al., 2025b ) : objects, ground literals of the state and goal (G- prefix), and ground action nodes with action–object edges carrying parameter slots. The lifted domain layer (top) is built from D alone: schemas, their precondition and effect occurrences, and predicate symbols. Binding edges (green, dashed) link an object’s slot on an applicable ground action to the precondition occurrences it satisfies; instantiation edges (orange, dotted) link each ground literal to its predicate symbol. Filled nodes are ground, outlined nodes lifted. The chain layer (magenta) is shown by one enables edge, one grounded deletes-true edge, and goal-relevance rings.
Figure 2: GADAR overview. (a) D , state, and goal become one graph. (b) The ladder: UNION (symbol one-hots), GADAR-BIND (structural features, lifted and binding layers), GADAR (chain layer). (c) GNN encoder. (d) GRU decoder: schema, then arguments, scored against node embeddings; nothing is sized by the number of schemas or objects.
Held out
Blocks
Gripper
Miconic
Visitall
Grid
Logistics
Spanner
Rovers
Mean
Training- scale
Rand
1.7
0.0
88.3
99.7
3.0
0.0
23.4
15.2
28.9
UNION
0.0
0.0
0.0
0.8
1.0
0.0
0.0
0.0
0.2
GADAR-BIND
2.5
0.0
0.0
16.0
3.1
0.0
0.0
0.0
2.7
GADAR
37.5
13.3
36.3
100.0
33.4
9.9
43.4
16.1
36.2
Test
Rand
0.2
0.0
1.1
6.0
0.0
0.0
13.8
5.2
3.3
UNION
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
Table 1: Zero-shot generalization, leave-one-domain-out: coverage (%) on each held-out domain, which is never seen in training. Training-scale: its small instances; Test: its test instances. Rand: uniform choice among applicable actions under the same executor. Mean: over the eight held-out domains.
Blocks
Gripper
Miconic
Visitall
Grid
Logistics
Spanner
Rovers
Mean
GABAR (published)
100.0
100.0
100.0
90.7
96.3
79.0
92.0
82.0
92.5
UNION
57.2
84.1
86.1
77.3
89.6
18.6
48.6
9.6
58.9
GADAR-BIND
43.4
95.4
67.0
78.6
64.9
0.3
40.0
8.2
49.7
GADAR
74.6
94.5
61.9
81.1
94.6
2.7
47.9
9.0
58.3
GADAR (single-domain)
95.5
100.0
41.2
54.0
87.5
3.1
85.4
7.4
59.3
Table 2: In-domain coverage (%) on the test instances. Ladder rows average the seven leave-one-out policies whose training set includes the domain; GADAR (single-domain) is trained on that domain alone. GABAR (published): per-domain GABAR as published, on its own test instances.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Domain
Problems
States
Blocks
200
4,216
Gripper
147
7,287
Miconic
228
5,453
Visitall
125
3,654
Grid
192
6,880
Logistics
156
6,732
Appendix
Table 3: Training data per domain: the small instances collected, and the labelled states taken from their plans.
Setting
Value
Encoder
Message-passing rounds L
9
Representation size
64
Attention heads
1
MLP layers per update block
2
Global node
enabled
Appendix
Table 4: Final hyperparameters, identical across all three systems except where noted.
Cross-domain generalization studies how machine-learning models retain reliable performance when data, sensors, populations, or tasks differ from training.
Classical planners can effectively solve very large deterministic MDPs represented in STRIPS or PDDL where states are sets of atoms over objects and relations, and lifted action schemas add or delete these atoms. This compact representation yields strong search heuristics and provides an ideal setting for structural generalization, since lifted relations and action schemas give rise to infinitely many domain instances. A central challenge is to learn these relations and action schemas from data, and recent approaches have addressed this problem using different types of observations. In this work, we develop a novel neural network architecture for learning action schemas from traces where states are fully observed but action arguments are unobserved. The problem is a simplification but an important step towards learning planning domains from sequences of images and action labels, and we aim to solve this simplification in a nearly perfect manner. The challenge lies in learning the action schemas while simultaneously identifying the action arguments from observed state changes. Our approach yields a robust differentiable component that can then be integrated into larger neuro-symbolic models. We evaluate the architecture on various planning domains, where the learned lifted action schemas must recover the ground-truth structure. Additionally, we report experiments on robustness to observation noise and on a variation related to slot-based dynamics models.
Learning per-domain generalizing policies is a key challenge in learning for planning. Standard approaches learn state-value functions represented as graph neural networks using supervised learning on optimal plans generated by a teacher planner. In this work, we advocate for learning Q-value functions instead. Such policies are drastically cheaper to evaluate for a given state, as they need to process only the current state rather than every successor. Surprisingly, vanilla supervised learning of Q-values performs poorly as it does not learn to distinguish between the actions taken and those not taken by the teacher. We address this by using regularization terms that enforce this distinction, resulting in Q-value policies that consistently outperform state-value policies across a range of 10 domains and are competitive with the planner LAMA-first.
Nicola J. Müller, Moritz Oster, Isabel Valera +2
German Research Center for Artificial Intelligence (DFKI), Saarbrücken, Germany · Saarland University, Saarland Informatics Campus, Saarbrücken, Germany · Center for European Research in Trusted Artificial Intelligence (CERTAIN)
We present GenePlan (GENeralized Evolutionary Planner), a novel framework that leverages large language model (LLM) assisted evolutionary algorithms to generate domain-dependent generalized planners for classical planning tasks described in PDDL. By casting generalized planning as an optimization problem, GenePlan iteratively evolves interpretable Python planners that minimize plan length across diverse problem instances. In empirical evaluation across six existing benchmark domains and two new domains, GenePlan achieved an average SAT score of 0.91, closely matching the performance of the state-of-the-art planners (SAT score 0.93), and significantly outperforming other LLM-based baselines such as chain-of-thought (CoT) prompting (average SAT score 0.64). The generated planners solve new instances rapidly (average 0.49 seconds per task) and at low cost (average $1.82 per domain using GPT-4o).
Andrew Murray, Danial Dervovic, Alberto Pozanco +1