Neurosymbolic Action Model Learning under Partial Observability
Adem Kikaj, Lennert De Smet, Giuseppe Marra, Luc De Raedt
cs.AI
Quick verdict
NeSyAM learns action models from partially observable traces; strong relevant-role recovery under occlusion, mixed full-model accuracy, no planning tests.
Summary
The paper addresses learning symbolic STRIPS action models (preconditions, add and delete effects) from visual planning traces when images are missing or only partially informative. Existing neurosymbolic methods, notably rosame-i, assume fully observable traces and, the authors argue, correspond to a mean-field variational approximation that enforces only local consistency between consecutive image predictions. The proposed NeSyAM embeds a probabilistic action model into a relational neurosymbolic Markov model: transitions are computed exactly under STRIPS constraints (including a non-standard rule that add effects apply only to false propositions, to curb reasoning shortcuts), and differentiable trajectory sampling supports either approximate maximum likelihood or a variational (ELBO) objective. Across six visual domains (Blocks World grid and synthesized, Gripper, Logistics, Hanoi, 8-Puzzle) and three observation regimes, NeSyAM recovers 90.00/90 relevant roles under full observability and 88.33-89.33 under occlusions, whereas rosame-i drops from 87.00 to 61.33-68.00; full-candidate recovery is mixed, with rosame-i better under full observability (183.33 vs 118.33 of 191). NeSyAM stays stable as Blocks World grows from 5 to 8 blocks but trains about 1.8x slower at 8 blocks; two trajectory samples suffice. Main limitations: false positives on not-involved predicates (identifiability), no downstream planning evaluation, three seeds without significance testing, and no code release.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Existing neurosymbolic action model learners, notably rosame-i, implicitly perform a mean-field variational approximation that relies on complete, fully observable visual traces and therefore fails under partial observability.
- NeSyAM exactly incorporates STRIPS semantics into the transitions of a relational neurosymbolic Markov model, enabling differentiable trajectory sampling for approximate maximum-likelihood or variational learning under partial observability.
- NeSyAM recovers all 90 relevant predicate roles under full observability and retains 88.33-89.33 under both occlusion regimes, while rosame-i drops from 87.00 to 61.33-68.00 (Table 1, three seeds).
- NeSyAM's recovery is stable as Blocks World grounding grows from 5 to 8 blocks (18/18 relevant roles), though training is about 1.8x slower than rosame-i at 8 blocks (Table 2).
- Remaining errors concentrate on false positives among not-involved predicates, an identifiability limitation of learning from positive-only traces (Section 7).
Strengths
Exact STRIPS-constrained probabilistic transitions
Section 4 derives a predicate-wise closed-form constrained transition (Eq. 17: pa/(pa+pn) when st=0; (pn+pp)/(pn+pp+pd) when st=1), replacing rosame-i's fuzzy relaxation, and builds the anti-reasoning-shortcut constraint (Eq. 8, add only on false propositions) into the model rather than into an auxiliary loss.
Unifying variational analysis of prior work
Section 5 recovers rosame-i's supervision, applicability and transition losses as the three terms of a mean-field ELBO (Eq. 21), giving a theoretical account of why image-wise local consistency fails when images underdetermine the state.
Robust relevant-role recovery under occlusion
Grounding stability and practical ablations
Candid identifiability analysis
Weaknesses
Mixed full-candidate profile underplayed at abstract level
No downstream planning evaluation
Thin statistical reporting
Single baseline; key details outside the main text
Cost and scale limits
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets lifted STRIPS action model learning from visual traces — images z1..zT, executed grounded actions, and a fully observed final-state description f — when intermediate images are missing or underdetermine the hidden symbolic states. Contributions: (1) NeSyAM, embedding Xi et al.'s (2024) probabilistic action model (pam) into Relational Neurosymbolic Markov Models (NeSyMMs, De Smet et al. 2025) with exact STRIPS-constrained transitions; (2) a variational reading casting rosame-i as a mean-field approximation; (3) experiments on six domains (grid and synthesized Blocks World, Gripper, Logistics, Hanoi, 8-Puzzle) x three observation regimes (O1 full, O2 random occlusion, O3 action occlusion). The abstract's claim that NeSyAM 'consistently recovers relevant parts of the true action model' is supported for the relevant-role subset Crel (90 of 90 candidates), but the full-candidate subset C tells a more mixed story that the abstract does not surface.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is NeSyAM?
NeSyAM is a neurosymbolic method that learns STRIPS action models (which predicates are preconditions, add effects, or delete effects of each action) from visual planning traces when intermediate images are missing or only partially informative. It combines a probabilistic action model with exact STRIPS-constrained state transitions inside a neurosymbolic Markov model and trains by sampling differentiable state trajectories.
How does NeSyAM differ from rosame-i?
rosame-i (Xi, Gould and Thiebaux, 2024) assumes complete, fully observable image traces and uses fuzzy, image-wise consistency losses that the authors show correspond to a mean-field variational approximation. NeSyAM instead computes exact constrained transitions, keeps sequential dependencies between hidden states in the model, and uses all images plus the final-state description in a single likelihood, which makes it more robust when images underdetermine the state.
What results does the paper report?
On six visual planning domains under three observation regimes (full, random occlusion, action occlusion), NeSyAM recovers 90 of 90 relevant predicate roles when fully observable and 88.33-89.33 under occlusion, while rosame-i drops from 87.00 to 61.33-68.00. On all 191 candidate action-predicate pairs, rosame-i is better under full observability (183.33 vs 118.33) and NeSyAM is better under random occlusion (124.67 vs 107.67).
Is code available for NeSyAM?
The paper does not state a public code release. Implementation details (perception and emission architectures, hyperparameters, the gradient estimator) are deferred to supplementary material, though the evaluation domains follow the public setup of Xi et al. (2024).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-2 are available only as captions; the HMM diagram and occlusion-mask visuals could not be inspected.
- Supplementary material (perception/emission architectures, hyperparameters, gradient derivation) was not part of the extracted text.
- Statistical significance cannot be independently verified from three-seed summaries alone.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.