VLA-Thinker: Boosting Vision-Language-Action Models through Thinking-with-Image Reasoning
VLA-Thinker boosts Vision-Language-Action models via dynamic image reasoning, achieving 97.5% success on LIBERO.
Key Findings
Methodology
Introduced a dynamic image reasoning framework treating visual perception as callable reasoning actions, trained via a two-stage pipeline: SFT cold-start and GRPO reinforcement learning.
Key Results
- Achieved 97.5% success rate on LIBERO, a 6.5% improvement over OpenVLA-OFT.
- Improved short, medium, and long-horizon tasks on RoboTwin 2.0 by 41%, 23.6%, and 18.1%, respectively.
- Ablation studies confirmed dynamic visual invocation significantly enhances task adaptability.
Significance
Addresses the static visual embedding limitation in traditional VLA models, enabling robust robotic manipulation in complex environments and advancing multimodal reasoning research.
Technical Contribution
First to model visual perception as dynamic reasoning actions, introducing GRPO-based reinforcement learning for optimizing long-horizon reasoning-action trajectories.
Novelty
VLA-Thinker pioneers dynamic image reasoning, breaking away from static visual embeddings and enhancing cross-modal interaction capabilities.
Limitations
- Only employs ZOOM-IN tool, leaving other visual tools unexplored.
- Model's generalization to real-world environments remains unverified.
Future Work
Future work could explore integrating diverse visual tools like 3D scanning or multi-view fusion and validate performance in real-world robotic settings.
AI Executive Summary
Traditional Vision-Language-Action models treat visual inputs as static context, limiting adaptability in long-horizon tasks. VLA-Thinker introduces a dynamic image reasoning framework that models visual perception as callable reasoning actions, optimized through a two-stage training pipeline.
Experiments show VLA-Thinker achieves a 97.5% success rate on the LIBERO benchmark, outperforming existing methods, and excels across diverse tasks in RoboTwin 2.0. Ablation studies further validate the effectiveness of dynamic visual invocation.
While currently limited to the ZOOM-IN tool, the framework offers extensibility for future multimodal reasoning research. Future directions include integrating additional visual tools and testing in real-world robotic environments.
Deep Analysis
Background
Vision-Language-Action models have advanced robotic manipulation but often rely on static visual embeddings, limiting adaptability in dynamic environments.
Core Problem
Static visual embeddings hinder long-horizon task performance, particularly in scenarios requiring iterative reasoning and dynamic adjustments.
Innovation
VLA-Thinker introduces dynamic image reasoning innovations:
- �� Models visual perception as callable reasoning actions.
- �� Employs a two-stage training pipeline: SFT cold-start and GRPO reinforcement learning.
- �� Utilizes ZOOM-IN tool for dynamic visual invocation.
Methodology
- �� SFT cold-start phase: Activates foundational reasoning capabilities using synthesized visual Chain-of-Thought data.
- �� GRPO reinforcement learning: Optimizes reasoning-action trajectories for task success.
- �� Dynamic visual invocation: Uses ZOOM-IN tool to retrieve task-relevant image details.
Experiments
Evaluated on LIBERO and RoboTwin 2.0 benchmarks, covering multi-task scenarios and bimanual collaborative tasks. Ablation studies confirm the impact of dynamic visual invocation.
Results
VLA-Thinker achieved 97.5% success on LIBERO, outperforming OpenVLA-OFT by 6.5%. Improved RoboTwin 2.0 short, medium, and long-horizon tasks by 41%, 23.6%, and 18.1%.
Applications
Applicable to robotic manipulation tasks such as home assistance, industrial automation, and complex object handling.
Limitations & Outlook
Currently limited to ZOOM-IN tool; broader visual tool integration remains unexplored. Real-world generalization requires further validation.
Plain Language Accessible to non-experts
Imagine a robot making coffee in a kitchen. Traditional robots observe the environment once and fail if they miss the coffee pot. VLA-Thinker acts like a smart assistant, dynamically rechecking the environment to locate the coffee pot and complete the task.
ELI14 Explained like you're 14
Imagine playing a robot game where the goal is making coffee. Regular robots only look at the kitchen once, and if they miss the coffee pot, they fail. VLA-Thinker is like a clever player—it can recheck the kitchen, find the pot, and win the game! Cool, right?
Glossary
Chain-of-Thought
A reasoning method that breaks tasks into intermediate steps for better decision-making.
Used to decompose long-horizon tasks into manageable steps.
GRPO
A reinforcement learning algorithm optimizing reasoning-action trajectories.
Trains the model to achieve task success.
ZOOM-IN
A visual tool for retrieving detailed images of specific regions.
Used for dynamic visual invocation in reasoning.
LIBERO
A benchmark dataset for multi-task robotic manipulation.
Evaluates task success rates.
RoboTwin 2.0
A simulation benchmark for bimanual collaborative tasks.
Tests model performance in complex scenarios.
Open Questions Unanswered questions from this research
- 1 How to integrate diverse visual tools for enhanced adaptability.
- 2 Validation of model generalization in real-world environments.
Applications
Immediate Applications
Home Robotics Assistant
Assists with household tasks like organizing items or cooking.
Industrial Automation
Handles complex object manipulation tasks in dynamic environments.
Long-term Vision
Intelligent Robotic Systems
Achieves fully autonomous dynamic decision-making and operation in complex environments.
Abstract
Vision-Language-Action (VLA) models have shown promising capabilities for embodied intelligence, but most existing approaches rely on text-based chain-of-thought reasoning where visual inputs are treated as static context. This limits the ability of the model to actively revisit the environment and resolve ambiguities during long-horizon tasks. We propose VLA-Thinker, a thinking-with-image reasoning framework that models perception as a dynamically invocable reasoning action. To train such a system, we introduce a two-stage training pipeline consisting of (1) an SFT cold-start phase with curated visual Chain-of-Thought data to activate structured reasoning and tool-use behaviors, and (2) GRPO-based reinforcement learning to align complete reasoning-action trajectories with task-level success. Extensive experiments on LIBERO and RoboTwin 2.0 benchmarks demonstrate that VLA-Thinker significantly improves manipulation performance, achieving 97.5% success rate on LIBERO and strong gains across long-horizon robotic tasks. Project and Codes: https://cywang735.github.io/VLA-Thinker/ .