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/ .
Figures & tables
Figure 1: Comparison between text-based CoT Reasoning (left) and Thinking-with-Image Reasoning (right) for VLA. Left: Conventional VLA reasoning models adopt a text-based Chain-of-Thought reasoning paradigm, treating visual inputs as static context, which fails to successfully grasp the target object. Right: Our proposed thinking-with-image framework models perception as a dynamically invocable reasoning action, enabling the model to call visual tools during intermediate reasoning steps and realize an interleaved perception–reasoning–action process, ultimately completing the manipulation task successfully.
Figure 2: The upper panel illustrates the main process of our proposed Thinking-with-Image framework . Language instructions and visual observations are encoded into a shared VLM, enabling interleaved reasoning and dynamic zoom-in perception before action generation. The lower panel presents the two-stage training strategy: (1) SFT cold-start to activate structured reasoning and tool-use behaviors, and (2) GRPO-based reinforcement learning to align multimodal reasoning–action trajectories with task-level objectives under sparse rewards.
Task Name
Steps
Horizon
Horizon Group
Short Horizon Tasks (112-130 steps)
lift_pot
112
Short
Average: 121 steps Count: 4 tasks
beat_block_hammer
113
Short
pick_dual_bottles
127
Short
place_phone_stand
130
Short
Medium Horizon Tasks (151-223 steps)
Table 1: RoboTwin 2.0 task classification based on planning horizon and required steps.
Model
LIBERO
Spatial
Object
Goal
Long
Avg
FlowVLA [ 68 ]
93.2
95.0
91.6
72.6
88.1
UnifiedVLA [ 32 ]
95.4
98.8
93.6
94.0
95.5
OpenVLA [ 28 ]
84.7
88.4
79.2
53.7
76.5
UniVLA [ 6 ]
96.5
96.8
95.6
92.0
95.2
CoT-VLA [ 65 ]
87.5
91.6
87.6
69.0
81.1
Table 2: Main results of different VLA models on LIBERO. All reported values denote the task success rate (SR, %) evaluated under 50 randomized initial conditions per task, averaged within each suite and across all suites. Bold numbers indicate the best performance within each suite.
Short Horizon Tasks (100-130 Steps)
Model
Lift Pot
Beat Hammer Block
Pick Dual Bottles
Place Phone Stand
Avg
π0 [ 3 ]
51.0
59.0
50.0
22.0
45.5
RDT [ 36 ]
45.0
22.0
18.0
13.0
24.5
π0 -FAST [ 42 ]
30.0
38.0
25.0
16.0
27.3
DeepThinkVLA [ 59 ]
62.0
73.0
61.0
24.0
55.0
OpenVLA-OFT [ 26 ]
10.1
28.1
29.7
17.1
21.3
Table 3: Main results of different VLA models on RoboTwin2.0. All reported values denote the task success rate (SR, %) evaluated under 100 randomized initial conditions per task. Bold numbers indicate the best performance within each task.
Method
Spatial
Object
Goal
Long
Avg.
OpenVLA-OFT [ 26 ]
91.6
95.3
90.6
86.5
91.0
VLA-Thinker-SFT
95.9
96.7
93.4
94.0
95.0
VLA-Thinker-GRPO
90.6
88.5
87.2
86.7
88.2
VLA-Thinker
98.7
99.0
95.2
96.9
97.5
Table 4: Ablation Study on training stages.
Figure 3: RL Training curves. (a) Task success reward steadily increases during GRPO training, demonstrating effective trajectory-level alignment under sparse rewards. (b) The average response length gradually decreases, indicating that the policy learns to invoke visual tools more selectively and reduce redundant reasoning.
Figure 4: Prompt template for training and inference.