aiqbench analysis
This paper introduces NS-VLA, a neuro-symbolic framework designed to overcome structural limitations in current Vision-Language-Action models for robotic manipulation. By integrating symbolic planning with neural backbones and optimizing reward granularity, the approach achieves superior zero-shot generalizability and robustness against data perturbations. The work represents a significant methodological advance in bridging the gap between structured reasoning and end-to-end learning for physical agents.
NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models
arXiv:2603.09542v2 Announce Type: replace Abstract: Vision-Language-Action (VLA) models are formulated to ground instructions in visual context and generate action sequences for robotic manipulation. Despite recent progress, VLA models still face structure-blind backbones, backbone-bound generalization, and flat single-objective optimization. To address these challenges, we propose a novel Neuro-Symbolic Vision-Language-Action (NS-VLA) framework. It introduces a Neuro-Symbolic Encoder for plan-constrained primitive inference, a Neuro-Symbolic Solver that conditions a backbone-agnostic policy on the active primitive, and Hierarchical Joint Policy Optimization with reward-granularity matching. Experiments on robotic manipulation benchmarks demonstrate that NS-VLA outperforms previous methods in both one-shot training and data-perturbed settings, while simultaneously exhibiting superior zero-shot generalizability and expanded exploration space. Our code is publicly available.
Abstract:Vision-Language-Action (VLA) models are formulated to ground instructions in visual context and generate action sequences for robotic manipulation. Despite recent progress, VLA models still face structure-blind backbones, backbone-bound generalization, and flat single-objective optimization. To address these challenges, we propose a novel Neuro-Symbolic Vision-Language-Action (NS-VLA) framework. It introduces a Neuro-Symbolic Encoder for plan-constrained primitive inference, a Neuro-Symbolic Solver that conditions a backbone-agnostic policy on the active primitive, and Hierarchical Joint Policy Optimization with reward-granularity matching. Experiments on robotic manipulation benchmarks demonstrate that NS-VLA outperforms previous methods in both one-shot training and data-perturbed settings, while simultaneously exhibiting superior zero-shot generalizability and expanded exploration space. Our code is publicly available.
32 pages, 10 figures, 2 tables. Major revision: updated author list and affiliations; revised methods, experiments, analysis, and appendices; added project page, code, model, and dataset links. Project page: this https URL
Robotics (cs.RO)
arXiv:2603.09542 [cs.RO]
(or arXiv:2603.09542v2 [cs.RO] for this version)
Full text under the publisher's open licence · 149 words
Related coverage
Vision-Predictor-Action: Towards Efficient Embodied AI via Neuro-Symbolic JEPA Predictors
arXiv · Oct 14, 2026
Towards Neuro-Symbolic Procedural Reasoning for Long-Horizon Vision-Language-Action Manipulation
arXiv · Sep 7, 2026
Academia is for Ambition — Alex Zhang, MIT
Latent Space · Oct 2, 2026
BRICS Draws a New Global Line: Full Text of New Delhi Declaration
slguardian.org · Sep 12, 2026