Paper deep dive
VLA-Thinker: Boosting Vision-Language-Action Models through Thinking-with-Image Reasoning
Chaoyang Wang, Wenrui Bao, Sicheng Gao, Bingxin Xu, Yu Tian, Yogesh S. Rawat, Yunhao Ge, Yuzhang Shang
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 98%
Last extracted: 3/22/2026, 5:10:14 AM
Summary
VLA-Thinker is a novel Vision-Language-Action (VLA) framework that introduces 'thinking-with-image' reasoning, allowing robotic models to dynamically invoke visual tools (like zoom-in) during the reasoning process. This approach moves beyond static text-based Chain-of-Thought by interleaving perception, reasoning, and action. The model is trained via a two-stage pipeline: SFT cold-start using synthesized CoT data, followed by GRPO-based reinforcement learning to align trajectories with task success. It achieves significant performance gains on LIBERO (97.5% success rate) and RoboTwin 2.0 benchmarks.
Entities (5)
Relation Signals (3)
VLA-Thinker â evaluatedon â LIBERO
confidence 100% ¡ We evaluate VLA-Thinker on two representative embodied intelligence benchmarks: LIBERO
VLA-Thinker â evaluatedon â RoboTwin 2.0
confidence 100% ¡ Experimental results demonstrate that VLA-Thinker achieves significant performance improvements on both benchmarks... RoboTwin 2.0
VLA-Thinker â usesalgorithm â GRPO
confidence 100% ¡ we employ Group Relative Policy Optimization (GRPO) to perform causal alignment
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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: this https URL .
Tags
Links
- Source: https://arxiv.org/abs/2603.14523v1
- Canonical: https://arxiv.org/abs/2603.14523v1
Trouble viewing inline? Open PDF directly â
Full Text
63,318 characters extracted from source content.
Expand or collapse full text
VLA-Thinker: Boosting Vision-Language-Action Models through Thinking-with-Image Reasoning Chaoyang Wang 1 Wenrui Bao 1 Sicheng Gao 2 Bingxin Xu 3 Yu Tian 1 Yogesh S Rawat 1 Yunhao Ge 4 Yuzhang Shang 1 * 1 University of Central Florida 2 University of WĂźrzburg 3 University of Southern California 4 NVIDIA Research Project&Codes:VLA-Thinker Abstract Enabling VisionâLanguageâAction (VLA) models to âthink before actingâ via Chain-of-Thought (CoT) reasoning has emerged as a promising direction for improving data ef- ficiency and decision robustness in embodied intelligence. However, existing CoT-enhanced VLA approaches remain constrained by a text-based paradigm: visual observations are encoded once as static context, while reasoning unfolds primarily in the language space. Such a design limits cross- modal interaction and prevents the model from actively re- visiting the environment to resolve ambiguities or recover from intermediate errors, particularly in long-horizon ma- nipulation tasks. To address these challenges, we propose VLA-Thinker, a thinking-with-image reasoning framework for embodied intelligence that aims to break away from text-based chain-of-thought reasoning by treating visual per- ception as an explicit component of the reasoning process. Unlike traditional VLA approaches that regard visual input as a one-shot observation, VLA-Thinker actively acquires task-relevant visual information through tool invocation dur- ing reasoning, thereby enabling an interleaved and cooper- ative perceptionâreasoningâaction process. Training such a system, however, presents unique challenges: the model must learn not only what to reason, but when and how to query visual information, and how to align complete reason- ingâaction trajectories with task success. To this end, we introduce a two-stage pipeline: (1) a SFT cold start phase using carefully curated visual CoT data to distill founda- tional reasoning capabilities and operation formats; and (2) the application of Group Relative Policy Optimization (GRPO) to causally align the complete reasoningâaction tra- jectories with desired task outcomes. Experimental results demonstrate that VLA-Thinker achieves significant perfor- mance improvements on both the LIBERO (97.5%) and the RoboTwin 2.0 (62.3%, 70.7%, and 64.6%). * Corresponding Author. 1. Introduction Vision-Language-Action (VLA) models have emerged as a promising paradigm in embodied intelligence, demonstrat- ing encouraging manipulation capabilities across a range of robotic tasks, such as stacking blocks, opening drawers, and organizing household objects. The prevailing approach is to learn a reactive end-to-end policy that directly maps high-level goals and perceptual inputs to low-level motor control commands [18,31,33,46,55â57,62]. However, this paradigm faces a critical bottleneck: learning such a holistic âperception-to-actionâ mapping is inherently chal- lenging and typically requires large amounts of high-quality demonstration data [37, 40]. To tackle the challenges of learning a direct perception- to-action mapping. A widely explored direction is to equip VLA models with the ability to âthink before actingâ, typi- cally instantiated through Chain-of-Thought (CoT) reason- ing [20,58,59,63]. Specifically, before producing actions, the model explicitly analyzes the task goal, current visual observations, object relationships, and potential subgoals, and generates a Chain-of-Thought (CoT) reasoning trace. Such reasoning enables the model to decompose high-level instructions into a sequence of executable intermediate deci- sions and to dynamically adjust its action strategy based on environmental feedback. However, existing VLA reasoning models remain con- strained by a text-based reasoning paradigm. In such ap- proaches, visual inputs are encoded once into static embed- dings and treated as fixed context throughout the reasoning process. Consequently, reasoning unfolds primarily in the language space, while perception becomes a passive, one- shot observation, as illustrated in Fig. 1 (left). This design departs significantly from human cognitive processes [2,45], where visual perception is active, iterative, and tightly cou- pled with reasoning. Humans dynamically revisit the en- vironment, selectively attend to task-relevant regions, and adapt visual focus when uncertainty arises. In contrast, static visual encoding limits a modelâs ability to resolve ambigui- arXiv:2603.14523v1 [cs.CV] 15 Mar 2026 <think> This is the initial frame. The stove (right-front) is off. The moka pot (center-front) and a pan (left-front) are on the counter. The stove knob (right-front, on stove) is accessible. To begin the global task, the stove must be turned on first. Turn on the stove. </think> <think> The robot arm is now positioned over the stove knob (right-front). The stove (right-front) remains off. The moka pot (center-front) and pan (left-front) are undisturbed. The knob (right-front, on stove) needs to be rotated to activatethe burner.Turn the stove knob to switch on the burner. </think> <think> The moka pot (on stove, right-front) is now situated on the active stove burner (right-front). The robot arm (top-center) is positioned above. The pan (left-front) is still on the counter. Both parts of the global task, turning on the stove and placing the moka pot on it, are complete. finish</think> <think> The robot arm has moved from its initial position and is now closer to the moka pot (center-front). The stove (right-front) remains on, and the pan (left-front) is undisturbed. The robot gripper is now ready to grasp the moka pot. The next action is to secure the moka pot. Pick up the moka pot.</think> <think> The robot arm has moved towards the stove knob (right-front). The stove (right-front) remains off, its burner is still. The moka pot (center-front) and pan (left-front) are undisturbed. The immediate goal is to fully tum the knob to activate the stove. Turn on the stove burner. </think> <think> This is the initial frame. The stove (right-front) is off, with its control knob (top-right of stove) accessible. The moka pot (center-front) and a pan (left-front) are on the counter. The first step is to activate the stove burner. Tum on the stove.</think> <think> The moka pot (right-front) is now situated on the active burner of the stove (right-front). The robot arm is positioned above the moka pot. The pan (left-front) is still on the counter. The global task to turn on the stove and put the moka pot on it is fulfilled.</think> <think>The cropped image shows that the robotic arm has already grasped the moka pot. The next logical action is to move the grasped moka pot onto the active burner of the stove. </think> <think> The robot arm is now positioned over the stove knob (top-right of stove). The stove (right front) is still off. The moka pot (center-front) and pan (left-front) are undisturbed. The next action is to operate the knob to tum on the stove. Turn on the stove</think> <think>The robotic arm has moved toward the stove (right-front). The moka pot (center-front) and the pan (left-front) remain undisturbed. I will now zoom in on theimage to obtain a clearer view.</think> <tool_call>ânameâ: âcrop_imageâ, âargumentsâ:"bbox_2d": [250, 80, 340, 200],</tool_call> <think>The cropped image shows that the stove knob is now directly beneath therobotâs end-effector. The immediate action is to interact with the knob.Rotate the stove knob to activate the burner.</think> <think>The stove burner (right-front) is now glowing red, indicating that it is on. I will now zoom in on the image to get a clearer view. </think> <tool_call>ânameâ: âcrop_imageâ, âargumentsâ: âbbox_2dâ: [270, 60, 390, 180], </tool_call> Text-based CoT Reasoning Thinking-with-Image Reasoning Figure 1. Comparison between text-based CoT Reasoning (left) and Thinking-with-Image Reasoning (right) for VLA. Left: Conven- tional 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 percep- tionâreasoningâaction process, ultimately completing the manipulation task successfully. ties, track subgoals, and recover from intermediate execution errors, particularly in long-horizon manipulation tasks. To address these challenges, we propose VLA-Thinker, a thinking-with-image reasoning framework for embodied intelligence. To the best of our knowledge, it is the first VLA model capable of thinkingâwithâimage reasoning. VLA- Thinker models perception as an explicit, dynamically in- vocable reasoning action. During reasoning process, the model can actively request task-relevant visual information (Relevant sub-image) through tool invocation, enabling per- ception to be interleaved with reasoning steps and action generation. This design transforms the traditional perceptionâ reasoningâaction pipeline into a tightly coupled and cooper- ative process, allowing the model to adapt its visual obser- vations based on evolving reasoning needs, as illustrated in Fig. 1 (right). Realizing such a perception-driven reasoning approach requires the model to learn not only structured reasoning patterns, but also when and how to query the envi- ronment effectively. To this end, we introduce a two-stage training strategy. First, a cold-start phase leverages carefully curated visual Chain-of-Thought data to distill foundational reasoning patterns and establish consistent operation formats for perception-driven reasoning. Second, we employ Group Relative Policy Optimization (GRPO) to perform causal alignment over complete reasoningâaction trajectories, en- couraging the model to generate effective perception queries and actions that jointly lead to task success. We evaluate VLA-Thinker on two representative embod- ied intelligence benchmarks: LIBERO [34] and RoboTwin 2.0 [9]. Experimental results demonstrate that VLA-Thinker achieves significant performance improvements on both benchmarks. In particular, VLA-Thinker attains a 97.5% success rate on the LIBERO benchmark, representing a 6.5% improvement over the backbone model OpenVLA-OFT [25], thereby validating the effectiveness of the proposed method. In summary, our contributions are threefold: â˘We introduce VLA-Thinker, the first VLA model capable of thinkingâwithâimage reasoning, which models visual perception as a dynamically invocable reasoning action, enabling Multimodal Embodied Chain-of-Thought. ⢠We propose a two-stage training framework combining SFT cold-start and GRPO-based trajectory-level align- ment, which stabilizes multimodal reasoning behaviors and effectively optimizes long-horizon reasoningâaction trajectories under sparse rewards. â˘Extensive experiments on multiple embodied benchmarks (LIBERO and RoboTwin 2.0) show the effectiveness of our proposed approach. Notably, VLA-Thinker achieves an average success rate of 97.5% on the LIBERO benchmark. 2. Method In this section, we present VLA-Thinker, the first thinking- with-image reasoning framework that tightly couples per- ception, reasoning, and action in embodied environments. Our method is built upon two key components. First, we reformulate VLA reasoning as an iterative multimodal in- terleaved process, where visual perception is treated as a dynamically invocable reasoning action rather than a static context. This design enables the model to actively query task- relevant visual evidence during intermediate reasoning steps and generate coherent reasoningâaction trajectories (Sec. 2.1). Second, we introduce a two-stage training strategy consisting of (1) a supervised fine-tuning (SFT) cold-start stage that activates structured reasoning and tool-use behav- iors using synthesized embodied Chain-of-Thought data, and (2) a trajectory-level reinforcement learning stage based on Group Relative Policy Optimization (GRPO), which aligns complete reasoningâaction trajectories with sparse task-level success signals (Sec. 2.2). Together, these components en- able VLA-Thinker to perform robust long-horizon reasoning and grounded action execution. 2.1. Problem Formulation We study Vision-Language-Action (VLA) reasoning [20,66] in embodied environments, where a model must generate action decisions by jointly reasoning over language instruc- tions and visual observations. Beyond conventional formula- tions that treat visual inputs as static context, we introduce a thinking-with-image reasoning paradigm that explicitly interleaves reasoning with visual perception. We formalize VLA thinking-with-image reasoning as an iterative multimodal interleaved reasoning process, in which perception is modeled as an explicit reasoning operation rather than a passive input. Given an initial language instructionT 0 and an initial visual observation setV 0 (e.g., egocentric RGB images), a VLA model iteratively produces a sequence of outputs: A k = f VLA T i ,C i ,V i k i=0 (1) where,T k denotes a textual reasoning step, representing the modelâs intermediate hypothesis or thought,C k denotes a perception invocation, specifying a visual tool invocation, V k denotes the visual evidence returned by executing the per- ception tool. A k denotes the action generated by the model. At each iteration, a controller (or parser) determines whether the model should: generate the next reasoning step and perception request(T k+1 ,C k+1 ), or terminate the rea- soning process and output an environment actionA. If a perception action is invoked, the corresponding visual tool is executed and returns new visual evidenceV k+1 , which is appended to the reasoning context and used to guide subse- quent reasoning and action generation. This process yields a multimodal reasoningâaction trajectory: Ď =T 1 ,C 1 ,V 1 ,T 2 ,C 2 ,V 2 ,...,T k ,A k ,(2) whereA k denotes the final environment action executed by the model. In this work, we consider one type of visual tool: ZOOM- IN, which is used to inspect fine-grained details within a specified region of the target image. The primary objective of this study is to validate the fundamental effectiveness of the interleaved perception-reasoning-action paradigm. There- fore, we employ the zoom-in mechanism as a representative instance to verify the end-to-end pipeline and demonstrate its potential for boosting VLA performance. We anticipate that this work will serve as a baseline, and we look forward to the community exploring more diverse and sophisticated visual tools in future follow-up research. Detailed protocols for the visual tool are provided in the Appendix. 2.2. Training Strategies We train VLA-Thinker using a two-stage training pipeline that first equips the model with foundational reasoning capabilities and subsequently aligns these capabilities with optimal task-level objectives, as illustrated in Fig. 2. Reasoning Activation via SFT Cold-Start. We begin with a cold-start supervised fine-tuning (Supervised Fine-Tuning, SFT) stage [8,53,54] to activate the modelâs foundational reasoning capabilities and tool-use behaviors. However, existing large-scale embodied intelligence datasets generally lack explicitly annotated Chain-of-Thought (CoT) reasoning trajectories, which substantially limits effective supervision of reasoning processes. To address this critical data gap, we leverage Qwen3-VL-30B-A3B-Instruct [1] to synthesize high-quality embodied CoT data. The generated reasoning trajectories include not only structured intermediate reasoning steps but also explicit modeling of valid and task-relevant tool invocation patterns. Specifically, we first identify semantically meaningful keyframes within each trajectory by detecting changes in the gripper state. Such state transitions typically correspond to subtask boundaries, enabling an effective decomposition of embodied tasks into hierarchical structures. For these keyframes, we employ Qwen3-VL-30B-A3B-Instruct to generate complete CoT annotations, including justified tool invocation and corre- sponding textual reasoning. For the remaining intermediate frames that are not selected as keyframes, we further use Qwen3-VL-30B-A3B-Instruct to generate pure textual CoT reasoning annotations, ensuring reasoning continuity throughout the entire trajectory. To guarantee the reliability VLM ActionHead ActionQuery Action Put the black bowl in the bottom drawer of the cab. <think>...The wine bottle (center-right) is nearby. I will zoom in on the image...</think> <think>...The next action is to close the drawer containing the bowl...</think> Task Instruction Image Image EmbeddingText Embedding Zoom-in Thinking-with- Image Reasoning Stage1.SFT ColdStart Tool Invocation Filtering SFT Data SFT Cold Start Stage2.Reinforcement Learning Group Relative Policy Optimization Update Policy ¡ Format Reward ¡Success Reward Tool Invocation 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 reasoningâaction trajectories with task-level objectives under sparse rewards. and consistency of the synthesized data, we enforce strict structured format validation (schema checks) and impose temporal consistency constraints on all generated annotations. Through this process, we construct a unified, clean, and high-quality embodied CoT dataset, which pro- vides a solid foundation for stable and effective SFT training. Learning Reasoning and Action via RL. After activating structured reasoning and valid tool-use behaviors via the SFT cold-start stage, we further optimize VLA-Thinker using reinforcement learning [30,47] to align complete reasoningâ action trajectories with task-level objectives. Different from conventional action-only policy learning [31,61], our goal is to jointly optimize reasoning steps, perception invocations, and environment actions under sparse and delayed rewards. We model VLA-Thinker as a stochastic policyĎ Î¸ over mul- timodal reasoningâaction trajectories. Given an instruction T 0 and initial visual observationV 0 , the policy generates a trajectory Ď =T 1 ,C 1 ,V 1 ,...,T n ,A n ,(3) whereT k denotes a textual reasoning step,C k a perception tool invocation,V k the returned visual sub-image content, and A n the final executable action. The reward functionR(Ď )is sparse and is assigned only at the end of a trajectory based on a verifiable task completion signalI success . No intermediate rewards are provided for the semantic correctness of the reasoning process itself. In addition, we introduce a small format regularization reward I format to prevent drift in the reasoning style. The reward function is defined as: R(Ď ) = Îą s ¡ I success + Îą f ¡ I format .(4) whereÎą s andÎą f are weighting coefficients. Here,I success = 1 if the task is successfully completed and0otherwise, while I format = 1if the CoT reasoning follows the correct format (<think></think><tool></tool>) and 0 otherwise. For each instruction and initial observation, we sample a group of M trajectories: Ď 1 ,Ď 2 ,...,Ď M âź Ď Î¸ .(5) Given their rewardsR(Ď i ) M i=1 , we compute the relative advantage for each trajectory as: A i = R(Ď i )â mean(R(Ď 1 ),R(Ď 2 ),...,R(Ď M )) std(R(Ď 1 ),R(Ď 2 ),...,R(Ď M )) (6) Following DeepSeek R1 [19], the training objective is de- fined as: J (θ) =E q,Ď i " 1 M M X i=1 min Ď Î¸ (Ď i |q) Ď Î¸ old (Ď i |q) A i , clip Ď Î¸ (Ď i |q) Ď Î¸ old (Ď i |q) , 1â Îľ, 1 + Îľ A i â βD KL (Ď Î¸ âĽĎ ref ) # (7) This relative formulation eliminates the need for an ex- plicit value function and substantially reduces variance when optimizing long-horizon reasoning trajectories with sparse feedback. By optimizingJ (θ), the VLA model is able to simultaneously enhance its reasoning capability and action execution capability, unifying both under the core objective of maximizing final task success. 2.3. Discussion The core philosophy of VLA-Thinker is to transition from a âpassive observationâ paradigm to an âactive percep- tionâreasoningâ paradigm. Traditional VLA models typ- ically treat vision as a static, one-shot input, thereby decou- pling perception from the subsequent multi-step reasoning process. In contrast, by modeling perception as a dynam- ically invocable reasoning action, our framework enables the model to actively revisit the environment to resolve am- biguities and recover from execution errors. The synergy between our two-stage training strategy is crucial to real- izing this philosophy. First, the SFT cold-start phase does more than teach the model to âverbalizeâ reasoning; it es- tablishes fundamental causal links between specific visual uncertainties and the necessity of tool invocation, injecting structured priors into the policy. Subsequently, GRPO re- inforces complete (thought, tool, action) trajectories using task-level success signals, effectively optimizing not only âhow to reasonâ but also âwhen to invoke.â Through this pro- cess, the model learns to balance reasoning cost against task success, ultimately learning when not to think to avoid redun- dant computation. Although this work primarily employs ZOOM-IN as the visual tool, our fundamental objective is to validate the effectiveness and feasibility of the inter- leaved reasoning paradigm within the VLA architecture. The zoom-in mechanism serves as a representative instance to instantiate the full pipeline and demonstrate its potential to enhance decision robustness in complex manipulation. We believe that the primary contribution of VLA-Thinker lies in this extensible framework rather than in any specific tool design. 3. Experiment 3.1. Experimental Setup Benchmarks. We evaluate VLA-Thinker on the LIBERO benchmark [35] and the RoboTwin 2.0 benchmark [9]. LIBERO is a language-guided manipulation benchmark designed for lifelong learning, covering diverse object types, task specifications, and environment settings. It consists of five task suites: LIBERO-Goal, LIBERO-Spatial, LIBERO-Object, LIBERO-Long (10 tasks, each with 50 expert demonstrations), and LIBERO-90 (containing 90 tasks for large-scale multi-task evaluation). We use the average Success Rate (SR) over 50 held-out test scenes per Table 1. RoboTwin 2.0 task classification based on planning horizon and required steps. Task NameStepsHorizonHorizon Group Short Horizon Tasks (112-130 steps) lift_pot112Short Average: 121 steps Count: 4 tasks beat_block_hammer113Short pick_dual_bottles127Short place_phone_stand130Short Medium Horizon Tasks (151-223 steps) move_can_pot151Medium Average: 176 steps Count: 4 tasks place_a2b_left155Medium place_empty_cup174Medium handover_mic223Medium Long Horizon Tasks (283-313 steps) handover_block283LongAverage: 298 steps Count: 2 tasksstack_bowls_two313Long Extra Long Horizon Tasks (466-637 steps) blocks_rank_rgb466Extra-LongAverage: 552 steps Count: 2 tasksput_bottles_dustbin637Extra-Long Overall StatisticsTotal: 12 tasks, Average: 256 steps task as the evaluation metric. RoboTwin2.0 is a simulation benchmark for bimanual manipulation, comprising 50 dual arm collaborative tasks and covering diverse robot morphologies and 731 object instances. The benchmark incorporates comprehensive domain randomization mecha- nisms (clutter, lighting, background, tabletop height, and language instructions), which enhance task diversity and improve sim to real generalization and transfer capability. During training and evaluation on RoboTwin2.0, we adopt domain randomized task settings and evaluate each task on 100 held out test scenarios. Specifically, we select 12 representative tasks and categorize them into four different temporal horizon levels based on their average execution steps, enabling a stratified and comprehensive evaluation. Tab. 1 summarizes the average number of steps for each task, as well as the step ranges corresponding to different horizon levels. Backbones. We adopt OpenVLA-OFT [26] as the base model. The model is built upon OpenVLA [28], adopting a vision encoder and LLaMA2-7B [52] as the backbone, and incorporating action chunking together with a parallel decoding design. This architecture provides high efficiency in online reinforcement learning scenarios that require frequent inference. To improve training and inference efficiency, we use only single view images, language instructions, and robot proprioceptive states as model inputs, while the official model additionally utilizes wrist camera images. Moreover, for the LIBERO tasks, we do not use robot proprioceptive states as inputs. In terms of model 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. Model LIBERO Spatial Object Goal Long Avg FlowVLA [68]93.295.091.672.688.1 UnifiedVLA [32]95.498.893.694.095.5 OpenVLA [27]84.788.479.253.776.5 UniVLA [6]96.596.895.692.095.2 CoT-VLA [65]87.591.687.669.081.1 WorldVLA [7]87.696.283.460.081.8 TraceVLA [67]84.685.275.154.174.8 MolmoAct [29]87.095.487.677.286.6 ThinkAct [21]88.391.487.170.984.4 PD-VLA [49]95.596.794.991.794.7 4D-VLA [64]88.995.290.979.188.6 SpatialVLA [43]88.289.978.655.578.1 Ď 0 [3]96.898.8 95.885.294.2 Ď 0 -FAST [42]96.496.888.660.285.5 NORA [22]92.295.489.474.687.9 SmolVLA [48]93.094.091.077.088.8 GR00T N1 [41]94.497.693.090.693.9 GraspVLA [12]-94.191.282.089.1 Seer â [51]---78.778.7 VLA-OS [17]87.096.592.766.085.6 Diffusion Policy â [10]78.392.568.350.572.4 OpenVLA-OFT91.695.390.686.591.0 VLA-Thinker (Ours) 98.799.0 95.2 96.9 97.5 â+7.1+3.7+4.6+10.4+6.5 architecture, we adopt only the parallel decoding and action chunk designs. Implementation Details. VLA-Thinker is initialized from the publicly available OpenVLA-OFT weights [26] and is trained on 8 NVIDIA H100 GPUs. During training and infer- ence, we only use single-view images, language instructions, and robot proprioceptive states as model inputs, whereas the official OpenVLA-OFT model [25] additionally incorporates wrist camera images. Besides, in the LIBERO, we donât in- clude robot proprioceptive states in model inputs. The model is trained using a two-stage pipeline. In the first stage, a cold-start supervised fine-tuning (SFT) is performed on our constructed embodied Chain-of-Thought (CoT) dataset. In the second stage, we further introduce online reinforcement learning (online RL) to align the generated CoT reason- ing with downstream action execution. This stage leverages outcome-based reward signals and is trained using the GRPO reinforcement learning algorithm. The batch size is set to 64 during the SFT stage and 128 during the RL stage. The learning rate is configured as1Ă 10 â5 for SFT and2Ă 10 â6 for RL, and both stages are optimized using the AdamW optimizer [38]. Overall, the complete training process takes approximately 3 days. Additional training details, includ- 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. Short Horizon Tasks (100-130 Steps) ModelLift PotBeat Hammer BlockPick Dual BottlesPlace Phone StandAvg Ď 0 [3]51.059.050.022.045.5 RDT [36]45.022.018.013.024.5 Ď 0 -FAST [42]30.038.025.016.027.3 DeepThinkVLA [59]62.073.061.024.055.0 OpenVLA-OFT [25]10.128.129.717.121.3 VLA-Thinker (Ours)64.882.565.436.662.3 â+54.7+54.4+35.7+19.5+41.0 Medium Horizon Tasks (150-230 Steps) ModelMove Can PotPlace A2B LeftPlace Empty CupHandover MicAvg Ď 0 [3]41.038.060.096.058.8 RDT [36]33.021.042.095.047.8 Ď 0 -FAST [42]34.036.054.083.051.8 DeepThinkVLA [59]52.038.083.088.065.3 OpenVLA-OFT [25]28.137.577.345.347.1 VLA-Thinker (Ours)61.039.192.789.970.7 â+32.9+1.6+15.3+44.6+23.6 Long (280-320 Steps) & Extra Long Horizon Tasks (450-650 Steps) ModelHandover BlockStack Bowls TwoBlocks Rank RgbPut Bottles DustbinAvg Ď 0 [3]39.053.045.036.043.3 RDT [36]26.042.017.026.027.8 Ď 0 -FAST [42]32.048.028.027.033.8 DeepThinkVLA [59]43.062.077.049.057.8 OpenVLA-OFT [25]33.140.670.242.246.5 VLA-Thinker (Ours)52.871.179.355.464.6 â+19.7+30.5+9.1+13.2+18.1 ing dataset statistics, hyperparameter configurations, and inference settings, are provided in Appendix. 3.2. Main Results LIBERO Benchmark Comparison. We evaluate VLA- Thinker on the LIBERO benchmark, which consists of four language-conditioned manipulation suites (Spatial, Object, Goal, and Long), covering diverse structured reasoning and long-horizon control scenarios. As shown in Tab. 2, VLA-Thinker achieves 98.7%, 99.0%, 95.2%, and 96.9% success rates on the four suites, respectively, yielding an overall average of 97.5%, which establishes a new state-of-the-art performance among all compared VLA models. Compared with the strong OpenVLA-OFT baseline (91.0% Avg.), our method achieves a substantial +6.5% overall improvement, with particularly pronounced gains on the Spatial (+7.1%) and Long (+10.4%) suites. These improvements indicate that explicitly modeling perception as a dynamically invocable reasoning action significantly enhances spatial grounding and long-horizon stability. The strong performance across all four suites suggests that integrating perception into the reasoning loop leads to more robust subgoal tracking, better ambiguity resolution, and im- proved action consistency under complex task specifications. RoboTwin2.0 Benchmark Comparison. We further evalu- ate VLA-Thinker on RoboTwin 2.0, a challenging dual-arm manipulation benchmark characterized by strong domain randomization and extended planning horizons. As sum- marized in Tab. 3, VLA-Thinker achieves 62.3% average success on short-horizon tasks (100â130 steps), outperform- Table 4. Ablation Study on training stages. MethodSpatial Object Goal Long Avg. OpenVLA-OFT [25]91.695.390.686.591.0 VLA-Thinker-SFT95.996.793.494.095.0 VLA-Thinker-GRPO90.688.587.286.788.2 VLA-Thinker98.799.0 95.2 96.9 97.5 ingĎ 0 (45.5%), DeepThinkVLA (55.0%), and OpenVLA- OFT (21.3%) by large margins. On medium-horizon tasks (150â230 steps), performance increases to 70.7%, exceeding DeepThinkVLA by over 5% and OpenVLA-OFT by more than 20%, demonstrating improved stability under moderate planning complexity. For long and extra-long horizon tasks (280â650 steps), VLA-Thinker achieves 64.6% average suc- cess, with notable gains on tasks such as Handover Block and Stack Bowls Two. Importantly, the relative performance ad- vantage becomes more significant as task horizon increases, suggesting that thinking-with-image reasoning effectively mitigates error accumulation in long reasoningâaction chains. By dynamically revisiting the environment and selectively querying task-relevant visual evidence, the model maintains coherent subgoal progression and exhibits stronger recov- ery capability when intermediate execution deviations occur. These results collectively validate that integrating active per- ception into the reasoning process is particularly beneficial for extended dual-arm coordination and complex temporal planning scenarios. 3.3. Ablation Study We conduct ablation studies to analyze (1) the contribution of thinking-with-image reasoning and (2) the effectiveness of the two-stage training pipeline. Ablation of Thinking-with-Image Reasoning. To isolate the effect of dynamic perception reasoning, we compare VLA-Thinker with OpenVLA-OFT. OpenVLA-OFT is an end-to-end visionâlanguageâaction policy model that directly predicts actions without explicit intermediate reasoning. As shown in Tab. 4, introducing the thinking- with-image reasoning mechanism improves the overall LIBERO performance from 91.0% to 97.5%. The gains are particularly pronounced in the Spatial and Long suites, where precise spatial grounding and consistent subgoal track- ing are critical. This comparison indicates that single-pass static visual encoding and direct action mapping are limited when handling fine-grained ambiguities and long-horizon tasks. In contrast, VLA-Thinker models perception as an explicitly invocable intermediate operation, enabling the model to query additional visual evidence through tool calls during decision-making. This leads to more grounded action selection under uncertainty. The performance improvements suggest that incorporating perception into the reasoning loop enhances policy robustness in complex scenarios and establishes a tighter perceptionâreasoningâaction coupling, rather than merely improving visual representation quality. Ablation of Training Pipeline. As shown in Tab. 4, we further investigate the contribution of each training stage by evaluating variants trained with only SFT cold-start or only GRPO reinforcement learning. The SFT-only model achieves 95.0% average success, demonstrating that struc- tured CoT supervision effectively activates reasoning format, tool-use patterns, and multimodal interaction behavior. How- ever, without trajectory-level reinforcement alignment, the model does not fully optimize reasoning for final task suc- cess. Conversely, directly applying GRPO without SFT ini- tialization results in severe performance degradation (88.2%), highlighting the instability of sparse-reward RL when struc- tured reasoning priors are absent. These findings confirm that SFT provides essential inductive biases and stabilizes reasoning behavior, while GRPO performs causal alignment over complete reasoningâaction trajectories. The combina- tion of both stages achieves the best performance (97.5%), demonstrating that reasoning activation and trajectory-level optimization are complementary and jointly indispensable for effectively training thinking-with-image VLA policies. 3.4. Training Curves Fig. 3 illustrates the RL training dynamics of VLA-Thinker. As shown in Fig. 3a, the task success reward exhibits a clear upward trend throughout training. Starting from an initial success level of approximately 0.82, the reward steadily in- creases and eventually converges near 0.96. This consistent improvement demonstrates that GRPO effectively aligns the multimodal reasoningâaction trajectories with the final task objective under sparse reward supervision. Importantly, the improvement is gradual rather than abrupt, indicating stable trajectory-level policy updates enabled by relative advantage normalization within sampled groups. Fig. 3b illustrates the trend of the average response length. We observe that as training progresses, the reasoning length gradually decreases. During the SFT cold-start stage, the model is trained mainly by imitating trajectories with complete reasoning processes. As a result, it tends to invoke tools frequently, even in rela- tively simple scenarios where additional visual queries are unnecessary, leading to a higher number of tool calls and comparatively redundant reasoning traces. After entering the RL stage, outcome-based policy optimization progressively reshapes the modelâs behavior. The model gradually learns to autonomously determine whether tool invocation is neces- sary based on task requirements. When critical information is missing or visual ambiguity exists, it actively requests additional visual evidence; when the current observation is sufficient for decision-making, it directly generates actions (a) Task Success Reward(b) Response Length 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. and avoids redundant tool calls. Eventually, the frequency of tool usage becomes more reasonable and stable, and the overall reasoning length correspondingly decreases. 4. Related Work VisionâLanguageâAction Models. VLA models unify perception, language understanding, and embodied ac- tion within a single framework. Early efforts such as SayCan [5] grounded large language models in robotic affordances, while Gato [44] and RT-1 [4] explored generalist multi-task transformers trained on large-scale demonstrations. PaLM-E [13] embedded continuous sensor modalities into a large language model for embodied reasoning. Subsequent work has broadened accessibility and scalability: the Open XâEmbodiment project [11] assembled a cross-embodiment dataset spanning dozens of robot types, and Octo [50] and OpenVLA [27] released open-source generalist policies efficiently fine-tunable to new robots. To improve action generation,Ď 0 [3] replaced autoregressive decoding with flow matching for high-frequency dexterous manipulation. Most recently, VLAs have been scaled to humanoid platforms: Helix [16] introduced a dual-system architecture for full upper-body humanoid control, GR00T N1 [41] combined a vision-language backbone with a diffusion-transformer action head trained on heterogeneous data. Unlike prior VLA models that primarily focus on architectural scaling or action representation improvements, VLA-Thinker redefines the role of perception. Instead of treating visual inputs as static context, we model perception as a dynamically invocable reasoning action, enabling interleaved perceptionâreasoningâaction. VLA Reasoning. Standard VLAs learn a direct observation- to-action mapping without intermediate reasoning, limiting generalization in complex, long-horizon tasks. To address this, recent work injects structured reasoning into VLAs through supervised fine-tuning (SFT): ECoT [14] introduces textual chain-of-thought reasoning with automatically gen- erated annotations covering plans, sub-tasks, and visual grounding before action prediction; CoT-VLA [65] instead generates subgoal images as a visual reasoning step leverag- ing action-free video data; RoboBrain [23] and Robix [15] further construct spatio-temporal thought-trace datasets with reinforced fine-tuning to strengthen causal reasoning. In the meantime, inspired by Large Reasoning Models, a par- allel line of work applies reinforcement learning (RL) to enhance embodied reasoning: Robot-R1 [24] and Embodied- R1 [60] use GRPO to reinforce VLM-based spatial reason- ing, with the former surpassing GPT-4o at only 7B param- eters and the latter achieving 87.5% zero-shot success on real-world robotic tasks. VLA-RL [39] and SimpleVLA- RL [30] apply online RL directly to auto-regressive VLA policies via trajectory-level formulations and scalable paral- lelization, attaining state-of-the-art performance on LIBERO and RoboTwin. While existing reasoning-enhanced VLAs rely on textual CoT supervision or action-level reinforce- ment learning, they largely remain text-based or optimize ac- tions independently of perception. VLA-Thinker integrates perception into the reasoning loop and performs trajectory- level GRPO alignment over complete multimodal reason- ingâaction sequences, enabling stable long-horizon reason- ing under sparse rewards. 5. Conclusion In this paper, we present VLA-Thinker, a thinking-with- image reasoning framework that integrates perception into the reasoning loop of VLA models. Unlike text-based CoT approaches that treat visual inputs as static con- text, our method models perception as a dynamically invocable reasoning action, enabling interleaved percep- tionâreasoningâaction trajectories. We further propose a two- stage training pipeline combining SFT-based reasoning acti- vation and GRPO-based trajectory-level alignment. Exten- sive experiments on LIBERO and RoboTwin 2.0 demonstrate that VLA-Thinker significantly outperforms strong baselines, achieving a 97.5% success rate on LIBERO. These results suggest that explicitly coupling perception with reasoning is crucial for robust and long-horizon embodied decision- making. References [1] Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., Ge, W., Guo, Z., Huang, Q., Huang, J., Huang, F., Hui, B., Jiang, S., Li, Z., Li, M., Li, M., Li, K., Lin, Z., Lin, J., Liu, X., Liu, J., Liu, C., Liu, Y., Liu, D., Liu, S., Lu, D., Luo, R., Lv, C., Men, R., Meng, L., Ren, X., Ren, X., Song, S., Sun, Y., Tang, J., Tu, J., Wan, J., Wang, P., Wang, P., Wang, Q., Wang, Y., Xie, T., Xu, Y., Xu, H., Xu, J., Yang, Z., Yang, M., Yang, J., Yang, A., Yu, B., Zhang, F., Zhang, H., Zhang, X., Zheng, B., Zhong, H., Zhou, J., Zhou, F., Zhou, J., Zhu, Y., Zhu, K.: Qwen3- vl technical report. arXiv preprint arXiv:2511.21631 (2025) 3 [2]Bajcsy, R., Aloimonos, Y., Tsotsos, J.K.: Revisiting active perception. Autonomous Robots 42(2), 177â196 (2018) 1 [3]Black, K., Brown, N., Driess, D., Esmail, A., Equi, M., Finn, C., Fusai, N., Groom, L., Hausman, K., Ichter, B., et al.:Ď0: A vision-language-action flow model for general robot control. corr, abs/2410.24164, 2024. doi: 10.48550. arXiv preprint arXiv:2410.24164 (2025) 6, 8 [4] Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Dabis, J., Finn, C., Gopalakrishnan, K., Hausman, K., Herzog, A., Hsu, J., Ibarz, J., Ichter, B., Irpan, A., Jackson, T., Jesmonth, S., Joshi, N.J., Julian, R., Kalashnikov, D., Kuang, Y., Leal, I., Lee, K.H., Levine, S., Lu, Y., Malla, U., Manjunath, D., Mordatch, I., Nachum, O., Parada, C., Peralta, J., Perez, E., Pertsch, K., Quiambao, J., Rao, K., Ryoo, M., Salazar, G., San- keti, P., Sayed, K., Singh, J., Sontakke, S., Stone, A., Tan, C., Tran, H., Vanhoucke, V., Vega, S., Vuong, Q., Xia, F., Xiao, T., Xu, P., Xu, S., Yu, T., Zitkovich, B.: Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817 (2023) 8 [5] Brohan, A., Chebotar, Y., Finn, C., Hausman, K., Her- zog, A., Ho, D., Ibarz, J., Irpan, A., Jang, E., Julian, R., et al.: Do as i can, not as i say: Grounding language in robotic affordances. In: Conference on robot learning. p. 287â318. PMLR (2023) 8 [6]Bu, Q., Yang, Y., Cai, J., Gao, S., Ren, G., Yao, M., Luo, P., Li, H.: Univla: Learning to act any- where with task-centric latent actions. arXiv preprint arXiv:2505.06111 (2025) 6 [7]Cen, J., Yu, C., Yuan, H., Jiang, Y., Huang, S., Guo, J., Li, X., Song, Y., Luo, H., Wang, F., Zhao, D., Chen, H.: Worldvla: Towards autoregressive action world model. arXiv preprint arXiv:2506.21539 (2025) 6 [8]Chen, D., Wang, C., SU, D., Xiao, X., Zhang, Z., Xiong, J., Li, Q., Shang, Y., Ka, S.: V-retrver: Evidence-driven agentic reasoning for universal mul- timodal retrieval. arXiv preprint arXiv:2602.06034 (2026) 3 [9]Chen, T., Chen, Z., Chen, B., Cai, Z., Liu, Y., Li, Z., Liang, Q., Lin, X., Ge, Y., Gu, Z., et al.: Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic ma- nipulation. arXiv preprint arXiv:2506.18088 (2025) 2, 5 [10] Chi, C., Xu, Z., Feng, S., Cousineau, E., Du, Y., Burchfiel, B., Tedrake, R., Song, S.: Diffusion pol- icy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research p. 02783649241273668 (2023) 6 [11] Collaboration, O.X.E., Padalkar, A., Pooley, A., Man- dlekar, A., Jain, A., Tung, A., Bewley, A., Herzog, A., Irpan, A., Khazatsky, A., Rai, A., Singh, A., Garg, A., Brohan, A., Raffin, A., Wahid, A., Burgess-Limerick, B., Kim, B., SchĂślkopf, B., Ichter, B., Lu, C., Xu, C., Finn, C., Xu, C., Chi, C., Huang, C., Chan, C., Pan, C., Fu, C., Devin, C., Driess, D., Pathak, D., Shah, D., BĂźchler, D., Kalashnikov, D., Sadigh, D., Johns, E., Ceola, F., Xia, F., Stulp, F., Zhou, G., Sukhatme, G.S., Salhotra, G., Yan, G., Schiavi, G., Kahn, G., Su, H., Fang, H.S., Shi, H., Amor, H.B., Christensen, H.I., Furuta, H., Walke, H., Fang, H., Mordatch, I., Radosavovic, I., Leal, I., Liang, J., Abou-Chakra, J., Kim, J., Peters, J., Schneider, J., Hsu, J., Bohg, J., Bing- ham, J., Wu, J., Wu, J., Luo, J., Gu, J., Tan, J., Oh, J., Malik, J., Booher, J., Tompson, J., Yang, J., Lim, J.J., SilvĂŠrio, J., Han, J., Rao, K., Pertsch, K., Hausman, K., Go, K., Gopalakrishnan, K., Goldberg, K., Byrne, K., Oslund, K., Kawaharazuka, K., Zhang, K., Rana, K., Srinivasan, K., Chen, L.Y., Pinto, L., FeiFei, L., Tan, L., Ott, L., Lee, L., Tomizuka, M., Spero, M., Du, M., Ahn, M., Zhang, M., Ding, M., Srirama, M.K., Sharma, M., Kim, M.J., Kanazawa, N., Hansen, N., Heess, N., Joshi, N.J., Suenderhauf, N., Palo, N.D., Shafiullah, N.M.M., Mees, O., Kroemer, O., Sanketi, P.R., Wohlhart, P., Xu, P., Sermanet, P., Sundaresan, P., Vuong, Q., Rafailov, R., Tian, R., Doshi, R., MartĂn- MartĂn, R., Mendonca, R., Shah, R., Hoque, R., Julian, R., Bustamante, S., Kirmani, S., Levine, S., Moore, S., Bahl, S., Dass, S., Sonawani, S., Song, S., Xu, S., Haldar, S., Adebola, S., Guist, S., Nasiriany, S., Schaal, S., Welker, S., Tian, S., Dasari, S., Belkhale, S., Osa, T., Harada, T., Matsushima, T., Xiao, T., Yu, T., Ding, T., Davchev, T., Zhao, T.Z., Armstrong, T., Darrell, T., Jain, V., Vanhoucke, V., Zhan, W., Zhou, W., Burgard, W., Chen, X., Wang, X., Zhu, X., Li, X., Lu, Y., Cheb- otar, Y., Zhou, Y., Zhu, Y., Xu, Y., Wang, Y., Bisk, Y., Cho, Y., Lee, Y., Cui, Y., Wu, Y.H., Tang, Y., Zhu, Y., Li, Y., Iwasawa, Y., Matsuo, Y., Xu, Z., Cui, Z.J.: Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In: 2024 IEEE International Conference on Robotics and Automation (ICRA). p. 6892â6903. IEEE (2024) 8 [12]Deng, S., Yan, M., Wei, S., Ma, H., Yang, Y., Chen, J., Zhang, Z., Yang, T., Zhang, X., Zhang, W., Cui, H., Zhang, Z., Wang, H.: Graspvla: a grasping foundation model pre-trained on billion-scale synthetic action data. arXiv preprint arXiv:2505.03233 (2025) 6 [13]Driess, D., Xia, F., Sajjadi, M.S., Lynch, C., Chowd- hery, A., Wahid, A., Tompson, J., Vuong, Q., Yu, T., Huang, W., et al.: Palm-e: An embodied multimodal language model (2023) 8 [14]Duan, Z., Zhang, Y., Geng, S., Liu, G., Boedecker, J., Lu, C.X.:Fast ecot:Efficient embodied chain-of-thought via thoughts reuse. arXiv preprint arXiv:2506.07639 (2025) 8 [15]Fang, H., Zhang, M., Dong, H., Li, W., Wang, Z., Zhang, Q., Tian, X., Hu, Y., Li, H.: Robix: A unified model for robot interaction, reasoning and planning. arXiv preprint arXiv:2509.01106 (2025) 8 [16]Figure:Helix:Avision-language-action modelforgeneralisthumanoidcontrol. https://w.figure.ai/news/helix (2025) 8 [17]Gao, C., Liu, Z., Chi, Z., Huang, J., Fei, X., Hou, Y., Zhang, Y., Lin, Y., Fang, Z., Jiang, Z., Shao, L.: Vla- os: Structuring and dissecting planning representations and paradigms in vision-language-action models. arXiv preprint arXiv:2506.17561 (2025) 6 [18]Guan, W., Hu, Q., Li, A., Cheng, J.: Efficient vision- language-action models for embodied manipulation: A systematic survey. arXiv preprint arXiv:2510.17111 (2025) 1 [19]Guo, D., Yang, D., Zhang, H., Song, J., Wang, P., Zhu, Q., Xu, R., Zhang, R., Ma, S., Bi, X., et al.: Deepseek- r1: Incentivizing reasoning capability in llms via rein- forcement learning. arXiv preprint arXiv:2501.12948 (2025) 4 [20]Guo, W., Lu, G., Deng, H., Wu, Z., Tang, Y., Wang, Z.: Vla-reasoner: Empowering vision-language-action models with reasoning via online monte carlo tree search. arXiv preprint arXiv:2509.22643 (2025) 1, 3 [21]Huang, C.P., Wu, Y.H., Chen, M.H., Wang, Y.C.F., Yang, F.E.: Thinkact: Vision-language-action reason- ing via reinforced visual latent planning. arXiv preprint arXiv:2507.16815 (2025) 6 [22]Hung, C.Y., Sun, Q., Hong, P., Zadeh, A., Li, C., Tan, U., Majumder, N., Poria, S., et al.: Nora: A small open-sourced generalist vision language action model for embodied tasks. arXiv preprint arXiv:2504.19854 (2025) 6 [23]Ji, Y., Tan, H., Shi, J., Hao, X., Zhang, Y., Zhang, H., Wang, P., Zhao, M., Mu, Y., An, P., et al.: Robobrain: A unified brain model for robotic manipulation from abstract to concrete. In: Proceedings of the Computer Vision and Pattern Recognition Conference. p. 1724â 1734 (2025) 8 [24]Kim, D., Park, S., Jang, H., Shin, J., Kim, J., Seo, Y.: Robot-r1: Reinforcement learning for en- hanced embodied reasoning in robotics. arXiv preprint arXiv:2506.00070 (2025) 8 [25]Kim, M.J., Finn, C., Liang, P.: Fine-tuning vision- language-action models: Optimizing speed and success. arXiv preprint arXiv:2502.19645 (2025) 2, 6, 7, 13 [26] Kim, M.J., Finn, C., Liang, P.: Fine-tuning vision- language-action models: Optimizing speed and success. arXiv preprint arXiv:2502.19645 (2025) 5, 6 [27]Kim, M.J., Pertsch, K., Karamcheti, S., Xiao, T., Bal- akrishna, A., Nair, S., Rafailov, R., Foster, E., Lam, G., Sanketi, P., Vuong, Q., Kollar, T., Burchfiel, B., Tedrake, R., Sadigh, D., Levine, S., Liang, P., Finn, C.: Openvla: An open-source vision-language-action model. In: The Conference on Robot Learning (CoRL) (2024) 6, 8 [28] Kim, M.J., Pertsch, K., Karamcheti, S., Xiao, T., Balakrishna, A., Nair, S., Rafailov, R., Foster, E., Lam, G., Sanketi, P., et al.: Openvla: An open- source vision-language-action model. arXiv preprint arXiv:2406.09246 (2024) 5 [29] Lee, J., Duan, J., Fang, H., Deng, Y., Liu, S., Li, B., Fang, B., Zhang, J., Wang, Y.R., Lee, S., Han, W., Pumacay, W., Wu, A., Hendrix, R., Farley, K., Vander- Bilt, E., Farhadi, A., Fox, D., Krishna, R.: Molmoact: Action reasoning models that can reason in space. arXiv preprint arXiv:2508.07917 (2025) 6 [30] Li, H., Zuo, Y., Yu, J., Zhang, Y., Yang, Z., Zhang, K., Zhu, X., Zhang, Y., Chen, T., Cui, G., et al.: Simplevla- rl: Scaling vla training via reinforcement learning. arXiv preprint arXiv:2509.09674 (2025) 4, 8 [31]Li, H., Ding, P., Suo, R., Wang, Y., Ge, Z., Zang, D., Yu, K., Sun, M., Zhang, H., Wang, D., et al.: Vla-rft: Vision-language-action reinforcement fine-tuning with verified rewards in world simulators. arXiv preprint arXiv:2510.00406 (2025) 1, 4 [32]Li, S., Gao, Y., Sadigh, D., Song, S.: Unified video action model. arXiv preprint arXiv:2503.00200 (2025) 6 [33] Li, X., He, X., Zhang, L., Wu, M., Li, X., Liu, Y.: A comprehensive survey on world models for embodied ai. arXiv preprint arXiv:2510.16732 (2025) 1 [34] Liu, B., Zhu, Y., Gao, C., Feng, Y., Liu, Q., Zhu, Y., Stone, P.: Libero: Benchmarking knowledge transfer for lifelong robot learning. Advances in Neural Infor- mation Processing Systems 36, 44776â44791 (2023) 2 [35]Liu, B., Zhu, Y., Gao, C., Feng, Y., Liu, Q., Zhu, Y., Stone, P.: Libero: Benchmarking knowledge transfer for lifelong robot learning. Advances in Neural Infor- mation Processing Systems 36, 44776â44791 (2023) 5, 13 [36]Liu, S., Wu, L., Li, B., Tan, H., Chen, H., Wang, Z., Xu, K., Su, H., Zhu, J.: Rdt-1b: a diffusion founda- tion model for bimanual manipulation. arXiv preprint arXiv:2410.07864 (2024) 6 [37]Liu, Y., Chen, W., Bai, Y., Liang, X., Li, G., Gao, W., Lin, L.: Aligning cyber space with physical world: A comprehensive survey on embodied ai. IEEE/ASME Transactions on Mechatronics (2025) 1 [38]Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) 6 [39]Lu, G., Guo, W., Zhang, C., Zhou, Y., Jiang, H., Gao, Z., Tang, Y., Wang, Z.: Vla-rl: Towards masterful and general robotic manipulation with scalable reinforce- ment learning. arXiv preprint arXiv:2505.18719 (2025) 8 [40] Ma, Y., Song, Z., Zhuang, Y., Hao, J., King, I.: A survey on vision-language-action models for embodied ai. arXiv preprint arXiv:2405.14093 (2024) 1 [41]NVIDIA, Bjorck, J., CastaĂąeda, F., Cherniadev, N., Da, X., Ding, R., Fan, L.J., Fang, Y., Fox, D., Hu, F., Huang, S., Jang, J., Jiang, Z., Kautz, J., Kundalia, K., Lao, L., Li, Z., Lin, Z., Lin, K., Liu, G., Llontop, E., Magne, L., Mandlekar, A., Narayan, A., Nasiriany, S., Reed, S., Tan, Y.L., Wang, G., Wang, Z., Wang, J., Wang, Q., Xiang, J., Xie, Y., Xu, Y., Xu, Z., Ye, S., Yu, Z., Zhang, A., Zhang, H., Zhao, Y., Zheng, R., Zhu, Y.: Gr00t n1: An open foundation model for generalist humanoid robots. arXiv preprint arXiv:2503.14734 (2025) 6, 8 [42] Pertsch, K., Stachowicz, K., Ichter, B., Driess, D., Nair, S., Vuong, Q., Mees, O., Finn, C., Levine, S.: Fast: Efficient action tokenization for vision-language-action models. arXiv preprint arXiv:2501.09747 (2025) 6 [43]Qu, D., Song, H., Chen, Q., Yao, Y., Ye, X., Ding, Y., Wang, Z., Gu, J., Zhao, B., Wang, D., et al.: Spatialvla: Exploring spatial representations for visual-language- action model. arXiv preprint arXiv:2501.15830 (2025) 6 [44]Reed, S., Zolna, K., Parisotto, E., Colmenarejo, S.G., Novikov, A., Barth-Maron, G., Gimenez, M., Sulsky, Y., Kay, J., Springenberg, J.T., et al.: A generalist agent. arXiv preprint arXiv:2205.06175 (2022) 8 [45]Sandini, G., Tistarelli, M.: Vision and space-variant sensing. In: Neural Networks for Perception, p. 398â 425. Elsevier (1992) 1 [46]Sapkota, R., Cao, Y., Roumeliotis, K.I., Karkee, M.: Vision-language-action (vla) models: Concepts, progress, applications and challenges. arXiv preprint arXiv:2505.04769 (2025) 1 [47] Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., Wu, Y., et al.: Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 (2024) 4, 13 [48] Shukor, M., Aubakirova, D., Capuano, F., Kooijmans, P., Palma, S., Adil Zouitine, M.A., Pascal, C., Russi, M., Marafioti, A., Alibert, S., Cord, M., Wolf, T., Ca- dene, R.: Smolvla: A vision-language-action model for affordable and efficient robotics. arXiv preprint arXiv:2506.01844 (2025) 6 [49]Song, W., Chen, J., Ding, P., Zhao, H., Zhao, W., Zhong, Z., Ge, Z., Ma, J., Li, H.: Accelerating vision- language-action model integrated with action chunking via parallel decoding. arXiv preprint arXiv:2503.02310 (2025) 6 [50]Team, O.M., Ghosh, D., Walke, H., Pertsch, K., Black, K., Mees, O., Dasari, S., Hejna, J., Kreiman, T., Xu, C., et al.: Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213 (2024) 8 [51]Tian, Y., Yang, S., Zeng, J., Wang, P., Lin, D., Dong, H., Pang, J.: Predictive inverse dynamics models are scal- able learners for robotic manipulation. arXiv preprint arXiv:2412.15109 (2025) 6 [52]Touvron, H., Martin, L., Stone, K., Albert, P., Alma- hairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhar- gava, P., Bhosale, S., et al.: Llama 2: Open foun- dation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) 5 [53] Wang, C., Feng, K., Chen, D., Wang, Z., Li, Z., Gao, S., Meng, M., Zhou, X., Zhang, M., Shang, Y., et al.: Adatooler-v: Adaptive tool-use for images and videos. arXiv preprint arXiv:2512.16918 (2025) 3 [54] Wang, C., He, Y., Zhou, Y., Wang, Y., Liu, J., Xia, P., Tu, Z., Bansal, M., Yao, H.: Knowing the answer isnât enough: Fixing reasoning path failures in lvlms. arXiv preprint arXiv:2512.06258 (2025) 3 [55]Wang, C., Zhang, Z., Meng, M., Zhou, X., Jiang, H.: Vision-ekipl: External knowledge-infused pol- icy learning for visual reasoning. arXiv preprint arXiv:2506.06856 (2025) 1 [56]Wang, C., Zhang, Z., Teng, L., Li, Z., Kan, S.: Tmcir: Token merge benefits composed image retrieval. arXiv preprint arXiv:2504.10995 (2025) [57]Wang, Y., Ding, P., Li, L., Cui, C., Ge, Z., Tong, X., Song, W., Zhao, H., Zhao, W., Hou, P., et al.: Vla-adapter: An effective paradigm for tiny- scale vision-language-action model. arXiv preprint arXiv:2509.09372 (2025) 1 [58]Ye, A., Zhang, Z., Wang, B., Wang, X., Zhang, D., Zhu, Z.: Vla-r1: Enhancing reasoning in vision-language- action models. arXiv preprint arXiv:2510.01623 (2025) 1 [59]Yin, C., Lin, Y., Xu, W., Tam, S., Zeng, X., Liu, Z., Yin, Z.: Deepthinkvla: Enhancing reasoning capabil- ity of vision-language-action models. arXiv preprint arXiv:2511.15669 (2025) 1, 6 [60]Yuan, Y., Cui, H., Huang, Y., Chen, Y., Ni, F., Dong, Z., Li, P., Zheng, Y., Hao, J.: Embodied-r1: Reinforced embodied reasoning for general robotic manipulation. arXiv preprint arXiv:2508.13998 (2025) 8 [61]Zang, H., Wei, M., Xu, S., Wu, Y., Guo, Z., Wang, Y., Lin, H., Shi, L., Xie, Y., Xu, Z., et al.: Rlinf-vla: A unified and efficient framework for vla+ rl training. arXiv preprint arXiv:2510.06710 (2025) 4 [62]Zhang, D., Sun, J., Hu, C., Wu, X., Yuan, Z., Zhou, R., Shen, F., Zhou, Q.: Pure vision language action (vla) models: A comprehensive survey. arXiv preprint arXiv:2509.19012 (2025) 1 [63] Zhang, D., Yuan, Z., Chen, Z., Liao, C.T., Chen, Y., Shen, F., Zhou, Q., Chua, T.S.: Reasoning-vla: A fast and general vision-language-action reason- ing model for autonomous driving. arXiv preprint arXiv:2511.19912 (2025) 1 [64]Zhang, J., Chen, Y., Xu, Y., Huang, Z., Zhou, Y., Yuan, Y.J., Cai, X., Huang, G., Quan, X., Xu, H., Zhang, L.: 4d-vla: Spatiotemporal vision-language-action pre- training with cross-scene calibration. arXiv preprint arXiv:2506.22242 (2025) 6 [65]Zhao, Q., Lu, Y., Kim, M.J., Fu, Z., Zhang, Z., Wu, Y., Li, Z., Ma, Q., Han, S., Finn, C., Handa, A., Liu, M.Y., Xiang, D., Wetzstein, G., Lin, T.Y.: Cot-vla: Visual chain-of-thought reasoning for vision-language-action models. In: Proceedings of the Computer Vision and Pattern Recognition Conference. p. 1702â1713 (2025) 6, 8 [66]Zhao, Q., Lu, Y., Kim, M.J., Fu, Z., Zhang, Z., Wu, Y., Li, Z., Ma, Q., Han, S., Finn, C., et al.: Cot-vla: Visual chain-of-thought reasoning for vision-language-action models. In: Proceedings of the Computer Vision and Pattern Recognition Conference. p. 1702â1713 (2025) 3 [67]Zheng, R., Liang, Y., Huang, S., Gao, J., DaumĂŠ I, H., Kolobov, A., Huang, F., Yang, J.: Tracevla: Vi- sual trace prompting enhances spatial-temporal aware- ness for generalist robotic policies. arXiv preprint arXiv:2412.10345 (2024) 6 [68]Zhong, Z., Yan, H., Li, J., Liu, X., Gong, X., Song, W., Chen, J., Li, H.: Flowvla: Thinking in motion with a visual chain of thought (2025) 6 A. Prompt Template Fig. 4 illustrates the prompt template for training and infer- ence of VLA-Thinker. B. Additional Implementation Details VLA-Thinker is initialized from the public OpenVLA-OFT weight [25]. Dataset Construction.Before performing Supervised Fine-Tuning (SFT), we construct two embodied Chain- of-Thought (CoT) datasets based on the public LIBERO demonstrations and Robotwin2.0 demonstrations, following the two-stage pipeline described in Section 2.2.This process generates 273,465 annotated keyframes and 215,784 annotated keyframes, respectively, which serve as the supervision data for the cold-start stage. Supervised Fine-Tuning (SFT). During the SFT stage, the model is trained for 100k steps using a batch size of 64 and a learning rate of1 Ă 10 â5 . We employ a hybrid attention mask that enables two complementary supervision modes within a single forward pass: CoT tokens are optimized autoregressively, while action tokens are supervised bidirectionally. Model parameters are optimized using a token-level cross-entropy loss. Reinforcement Learning (RL). In the reinforcement learn- ing stage, we adopt Group Relative Policy Optimization (GRPO) [47]. Each trajectory receives a sparse task-success reward, complemented by a small format-regularization re- ward to maintain the quality and consistency of generated CoT reasoning. Policy optimization is performed with a mini-batch size of 128, a low clip ratioÎľ = 0.2, and a high clip ratioÎľ = 0.28. Additionally, a KL penalty relative to the SFT reference model is introduced to mitigate catastrophic forgetting during policy updates. Infrastructure and Inference. Training is conducted on 8 NVIDIA H100 GPUs. During inference, we adopt greedy decoding for both reasoning and action tokens. C. Inference Speed We evaluate the inference efficiency of VLA-Thinker in comparison with the end-to-end OpenVLA-OFT [25] on the LIBERO benchmark [35] using an H100 GPU. On aver- age, VLA-Thinker requires 19% more execution time than OpenVLA-OFT, primarily due to its autoregressive reason- ing process. Despite this moderate increase in inference time, the proposed embodied reasoning mechanismâserving as a form of test-time scalingâsubstantially improves down- stream task performance. Specifically, VLA-Thinker consis- tently outperforms OpenVLA across all four LIBERO task categories, achieving success rate improvements of 7.1% on spatial tasks, 3.7% on object tasks, 4.6% on goal tasks, and 10.4% on long-horizon tasks. These results indicate that the additional computational overhead introduced by reasoning is well justified by the resulting performance gains, highlight- ing the effectiveness of embodied reasoning for enhancing robotic manipulation capabilities. D. Limitations and Future Works Although VLA-Thinker demonstrates strong performance on both LIBERO and RoboTwin 2.0 benchmarks, several limitations remain. First, the current framework employs only a single visual tool (ZOOM-IN) to validate the effec- tiveness of thinking-with-image reasoning. While sufficient to demonstrate the proposed paradigm, more diverse per- ception tools (e.g., object grounding, segmentation, or web- search) may further enhance reasoning capability in complex environments. Second, since VLA-Thinker is built upon pretrained multimodal large language models (MLLMs), it inevitably inherits their inherent limitations, particularly the issue of hallucination in visual or spatial reasoning. This may cause the generated actions to reference incorrect object attributes or spatial relationships, thereby affecting the sub- sequent execution process. Future progress in mechanisms for mitigating MLLM hallucinations could further improve the robustness and reliability of the system for real-world deployment. Prompt Template for Training and Inference You are an advanced robotic intelligence agent. # Tools You may call one functions to assist with the user query. You are provided with function signatures within <tools></tools> XML tags: <tools> "type": "function", "function": "name": "crop_image", "description": "Zoom in on the image based on the bounding box coordinates.", "parameters": "type": "object", "properties": "bbox_2d": "type": "array", "description": "coordinates for bounding box of thearea you want to zoom in. minimum value is 0 and maximum value is the width/height of the image.", "items": "type": "number", "target_image": "type": "number", "description": "The index of the image to crop. Index from 1 to the number of images. Choose 1 to operate on original image.", "required": ["bbox_2d", "target_image"] </tools> For each function call, return a jsonobject with function name and arguments within <tool></tool> XML tags: <tool> "name": <function-name>, "arguments": <args-json-object> </tool> Understand the given visual information and the task instruction. Determine if it is beneficial to employ the given visual operations (tools). Place your text reasoning process within the <think> </think> tags. Place any function calls within the <tool></tool> tags. Figure 4. Prompt template for training and inference.