Robots that read intent: logic-consistency bias in attention cuts contradictions by a third

Proactive human-robot collaboration needs systems to infer not only the current action but the hierarchical intentions governing behavior; this work extends action recognition from flat classification to four levels: atomic actions, composite activities, low-level intentions, and high-level intentions. Prior deep learning approaches focused predominantly on flat action classification and did not capture the structural dependencies linking kinematic motion to high-level goals. Evaluation uses the synthesized NTU-Hierarchy dataset, a four-level ontological extension of NTU RGB+D 120 that maps existing sequences into 119,383 hierarchical episodes spanning 85 actions, 16 activities, 8 low-level intentions, and 4 high-level intentions.

The hierarchy is not stitched on after the fact: a human reasoning graph enters self-attention through learnable topological biases, and a differentiable first-order logic loss acts as a structural regularizer enforcing ontological consistency across the four levels. Adding the first-order logic loss to the baseline improves logic consistency by 3.26 percentage points to 96.67%; in the consistency-maximizing configuration ($\lambda$ = 0.5), logic consistency reaches 94.59% and ontological contradictions drop by 33% while retaining 89.90% top-1 action accuracy. In the base configuration the neuro-symbolic graph transformer achieves 91.65% top-1 action accuracy, and strict isolation between label generation and first-order logic constraints prevents circular supervision.

The learned reasoning-graph weights provide interpretable reasoning evidence, with transition edges strongly exploited at $\beta$ trans approximately +2.6 and no explicit supervision. The human reasoning graph biases establish a tunable trade-off between accuracy and structural coherence. This weight behavior comes from the learned results in the same NTU-Hierarchy evaluation.

The method targets proactive human-robot collaboration settings that require hierarchical intention inference, and its evaluation scope is the synthesized NTU-Hierarchy dataset, a four-level ontological extension of NTU RGB+D 120 with 119,383 hierarchical episodes. The material states no next experiment or deployment plan, so the supported scope stays within that dataset and four-level ontology.

The material states that strict isolation between label generation and first-order logic constraints prevents circular supervision, but reports no real-world or cross-dataset validation; accuracy falls to 89.90% in the consistency-maximizing configuration, indicating a trade-off between accuracy and structural coherence.