Research map / Vision-language and multimodal / Robot action and navigation
NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NS-VLA, a neuro-symbolic VLA that infers a monotone plan of primitives with a pointer, conditions a chunked action backbone on the active primitive, and trains with hierarchical joint policy optimization using granularity-matched rewards.
Problem
VLA models are structure-blind, generalize poorly through the backbone to unseen instructions and data-scarce settings, and use flat trajectory-level RL that mismatches hierarchical policy structure.
Approach
A plan-constrained neuro-symbolic encoder yields a monotone primitive pointer that conditions a backbone-agnostic chunked action policy, and a hierarchical RL scheme gives sparse milestone rewards to the high-level primitive selector and dense per-chunk rewards to the low-level controller.
Key results
- On 1-shot LIBERO, NS-VLA reaches 69.1 average SR versus 65.3 for VLA-Adapter (second best).
- On LIBERO-Plus perturbed evaluation, NS-VLA reaches 49.8 average SR versus 28.0 for the second-best method, with robustness 72.0.
- On CALVIN ABC->D, NS-VLA reports 91.2 5-task SR and 4.72 average chain length versus 80.0 and 4.50 for VLA-Adapter.
Surprising findings
- Adding NS-VLA to VLA-Adapter yields only +3.8 points, the smallest gain among backbones, which the authors attribute to reduced headroom.
Limitations
- Real-world evaluation is limited to 100 trials on a single robot arm with three task categories.
- Propositions are stated with proofs deferred; the optimality claim relies on the shaping argument and an assumed bound on KL at optimum.
Experimental setup
Datasets: LIBERO LIBERO-Plus LIBERO-90 CALVIN ABC->D
Benchmarks: LIBERO LIBERO-Plus LIBERO-90 CALVIN ABC->D
Metrics: success rate robustness ratio plan-stage completion rate average completed chain length
Baselines: WorldVLA pi0 Diffusion Policy OpenVLA NORA UniVLA OpenVLA-OFT VLA-RL pi0-Fast VLA-Adapter SimpleVLA-RL RIPT-VLA
Builds on: GRPO OpenVLA-OFT VLA-Adapter potential-based reward shaping
Closest papers by meaning
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.93)
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.93)
- VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.92)
- FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies (NeurIPS 2026 · similarity 0.92)
- CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models (CVPR 2025 · similarity 0.92)
- Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.