[논문 리뷰] The Price Is Not Right: Neuro-Symbolic Methods Outperform VLAs on Structured Long-Horizon Manipulation Tasks with Significantly Lower Energy Consumption

Timothy R. Duggan, Pierrick Lorang외 2명|arXiv (Cornell University)|2026. 02. 22.

Multimodal Machine Learning Applications인용 수 0

한 줄 요약

논문은 구조화된 Towers of Hanoi 작업에서 신-기호 계획-제어 모델과 미세조정된 Vision-Language-Action (VLA) 모델을 비교하여 신-기호 접근법이 더 높은 작업 성공률과 현저히 낮은 에너지 사용을 달성함을 보이고 4블록 변형으로의 일반화를 포함합니다.

ABSTRACT

Vision-Language-Action (VLA) models have recently been proposed as a pathway toward generalist robotic policies capable of interpreting natural language and visual inputs to generate manipulation actions. However, their effectiveness and efficiency on structured, long-horizon manipulation tasks remain unclear. In this work, we present a head-to-head empirical comparison between a fine-tuned open-weight VLA model π0 and a neuro-symbolic architecture that combines PDDL-based symbolic planning with learned low-level control. We evaluate both approaches on structured variants of the Towers of Hanoi manipulation task in simulation while measuring both task performance and energy consumption during training and execution. On the 3-block task, the neuro-symbolic model achieves 95% success compared to 34% for the best-performing VLA. The neuro-symbolic model also generalizes to an unseen 4-block variant (78% success), whereas both VLAs fail to complete the task. During training, VLA fine-tuning consumes nearly two orders of magnitude more energy than the neuro-symbolic approach. These results highlight important trade-offs between end-to-end foundation-model approaches and structured reasoning architectures for long-horizon robotic manipulation, emphasizing the role of explicit symbolic structure in improving reliability, data efficiency, and energy efficiency. Code and models are available at https://price-is-not-right.github.io

연구 동기 및 목표

• Assess task performance and energy efficiency of neuro-symbolic vs. Vision-Language-Action models on structured, long-horizon manipulation tasks.
• Evaluate generalization to unseen task variants (e.g., 4-block Towers of Hanoi).
• Quantify training/inference energy consumption for both architectural paradigms.
• Analyze the impact of explicit symbolic structure on reliability and data efficiency.

제안 방법

• Head-to-head empirical comparison between a fine-tuned open-weight VLA model (π0) and a neuro-symbolic architecture combining PDDL-based symbolic planning with diffusion-based low-level control.
• Evaluate on simulated Towers of Hanoi variants (3-block and 4-block) in Robosuite; measure task success, advancement, and energy consumption during training and execution.
• Two VLA configurations tested: End-to-End (E2E-VLA) and Planner-Guided (PG-VLA); NSM uses symbolic planning plus neural skills learned from demonstrations.
• NSM abstracts symbolic operators from demonstrations via a minimal bisimulation of a learned graph, then solves with a classical planner (PDDL); low-level policies are diffusion-based and operate on relative end-effector poses.]

실험 결과

연구 질문

• RQ1 Does a neuro-symbolic architecture outperform fine-tuned VLAs on structured long-horizon manipulation tasks in terms of task success and generalization to unseen configurations?
• RQ2 What are the relative energy costs of training/fine-tuning and inference for NSM vs. VLA models?
• RQ3 Can explicit symbolic planning improve reliability and data efficiency over end-to-end VLA approaches on multi-step tasks like Towers of Hanoi?
• RQ4 How well do VLAs generalize to higher-block variants (e.g., 4-block) compared to NSM?

주요 결과

• On the 3-block Towers of Hanoi, NSM achieves 95% success versus 34% for the best-performing VLA.
• NSM generalizes to the unseen 4-block variant with 78% success; both VLAs fail to complete the 4-block task.
• Training energy: NSM requires ~0.65–0.85 MJ total energy vs. ~64–68 MJ for VLAs, i.e., nearly two orders of magnitude lower for NSM.
• Inference: VLAs consume substantially more energy overall due to GPU-backed inference; NSM uses no GPU during inference.
• NSM achieves near-perfect performance on the 3-block task with significantly faster per-episode times than VLAs (e.g., Individual Move: NSM 6.3 s vs. E2E-VLA 13.8 s).
• VLM-based planners (GPT-5, Qwen, PaLI-Gemma) yield limited planning accuracy and high per-query energy, underscoring the instability and cost of VLM-driven planning.