cs.ROMar 1, 2026

DAM-VLA: A Dynamic Action Model-Based Vision-Language-Action Framework for Robot Manipulation

Authors: Xiongfeng Peng, Jiaqian Yu, Dingzhe Li, Yixiang Jin, Lu Xu, Yamin Mao, Chao Zhang, Weiming Li, +3 more

Organizations: Advanced Research Lab, Samsung R&D Institute China-Beijing (SRCB), China · Samsung AI Center, DS Division, South Korea · Hanyang University ERICA, South Korea

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

In dynamic environments such as warehouses, hospitals, and homes, robots must seamlessly transition between gross motion and precise manipulations to complete complex tasks. However, current Vision-Language-Action (VLA) frameworks, largely adapted from pre-trained Vision-Language Models (VLMs), often struggle to reconcile general task adaptability with the specialized precision required for intricate manipulation. To address this challenge, we propose DAM-VLA, a dynamic action model-based VLA framework. DAM-VLA integrates VLM reasoning with diffusion-based action models specialized for arm and gripper control. Specifically, it introduces (i) an action routing mechanism, using task-specific visual and linguistic cues to select appropriate action models (e.g., arm movement or gripper manipulation), (ii) a dynamic action model that fuses high-level VLM cognition with low-level visual features to predict actions, and (iii) a dual-scale action weighting mechanism that enables dynamic coordination between the arm-movement and gripper-manipulation models. Across extensive evaluations, DAM-VLA achieves superior success rates compared to state-of-the-art VLA methods in simulated (SIMPLER, FurnitureBench) and real-world settings, showing robust generalization from standard pick-and-place to demanding long-horizon and contact-rich tasks.

Figures & tables

Fig. 2: We identify three distinctions between the arm movement and the gripper manipulation using the task of placing a carrot on a plate as an illustrative example: Path Constrains, Visual Attention, and Dataset Representation.
Fig. 3: The architecture of our DAM-VLA. Given an RGB image observation and a task description, the model predicts a sequence of temporal actions. The process consists of three key components: 1) a vision-language model that encodes observation into visual, class and register tokens, and integrates visual tokens with a set of linguistic tokens, and produces the cognition and reasoning latents; 2) an action routing module that generates a weight and feeds it into the dynamic action model; 3) a dynamic action model that dynamically executes different action models by combining the low-level class token or register token from the vision model with the high-level cognition latent from the VLM to predict an action sequence.
Fig. 4: Illustration of the dual-scale action weighting mechanism. The trajectory weight highlights critical manipulation phases via asymmetrical Gaussian distributions. Within each predicted chunk, the action chunk weight applies exponential decay to prioritize immediate temporal accuracy. The final weight integrates both scales to guide model supervision.
Method / Google(VA)Success Rates on Different TasksAvg
PCCMNOCDODPA
RT-1 [ 3 ]90%46%35%3%44%
RT-1-X [ 44 ]49%33%29%10%30%
RT-2-X [ 44 ]82%79%35%21%54%
Octo-Base [ 37 ]1%4%1%0%1%
RoboVLM [ 45 ]76%60%11%--
TABLE I: Comparison of success rates between our method and existing VLA methods on the Google robot in Variant Aggregation (VA) setting of the SIMPLER simulated evaluation across four tasks. (PCC: Pick Coke Can, MN: Move Near, OCD: Open/Close Drawer, ODPA: Open Drawer and Place Apple, Avg: Average)
Method / Google(VM)Success Rates on Different TasksAvg
PCCMNOCDODPA
RT-1 [ 3 ]87%39%72%8%51%
RT-1-X [ 44 ]59%33%56%17%41%
RT-2-X [ 44 ]79%78%25%4%46%
Octo-Base [ 37 ]18%4%24%0%11%
OpenVLA [ 4 ]14%51%48%0%28%
TABLE II: Comparison of success rates between our method and existing VLA methods on the Google robot in Visual Matching (VM) setting of the SIMPLER simulated evaluation across four tasks. (PCC: Pick Coke Can, MN: Move Near, OCD: Open/Close Drawer; ODPA: Open Drawer and Place Apple, Avg: Average; + : open-pi-zero , which is trained on Fractal dataset.)
Method / WidowX(VM)Success Rates on Different TasksAvg
SoTCoPGoYEiB
RT-1-X [ 44 ]4%8%0%0%3%
Octo-Base [ 37 ]7%13%0%44%16%
Octo-Small [ 37 ]47%8%1%51%27%
OpenVLA [ 4 ]4%0%0%4%2%
ECOT [ 7 ]4%8%0%0%3%
TABLE III: Comparison of success rates between our method and existing VLA methods on the WidowX robot in the Visual Matching (VM) setting of the SIMPLER simulated evaluation across four tasks. (SoT: Put Spoon on Towel, CoP: Put Carrot on Plater; GoY: Stack Green Block on Yellow Block; EiB: Put Eggplant in Yellow Basket; Avg: Average; + : open-pi-zero , which is trained on BridgeDataV2.)
Fig. 5: The entire process of the “One-Leg” assembly task in the FurnitureBench environment.
Method / FurnitureBenchSuccess Rates at Each Step
12345
OpenVLA [ 4 ]96%94%78%53%29%
CogACT [ 12 ]98%96%96%56%42%
DAM-VLA(Ours)100%100%100%62%56%
TABLE IV: The success rate of each step of the “One-Leg” assembly task in FurnitureBench is compared with the existing OpenVLA and CogACT methods. The fifth step represents the final success rate of the task.
Fig. 6: The evaluation encompasses both in-distribution and out-of-distribution scenarios. The in-distribution setting includes variations in object positions and lighting conditions consistent with the training data, while the out-of-distribution setting introduces novel backgrounds, unseen objects, and visual distractors absent during training.
Method / Real-worldSuccess Rates on Different Scenarios
IDOODAvg
CogACT65.7%60.0%62.9%
DAM-VLA (Ours)91.4%82.2%86.8%
TABLE V: The success rates of our method compared with CogACT on the pick-and-place task in the real-world evaluation. ID: In-Distribution, OOD: Out-of-Distribution.
ComponentsSuccess Rates
VTACWTWDAMDLGoogle (VM)Google (VA)WidowX (VM)Avg
-----64%61%50%58%
✓✓---78%68%53%66%
✓✓✓--76%63%51%63%
✓✓✓✓-82%72%43%66%
✓✓✓✓✓83%81%71%78%
TABLE VI: An ablation study is conducted on the WidowX robot in Visual Matching (VM) setting and the Google robot in both VM and Variant Aggregation (VA) settings of the SIMPLER environment.

Explore similar work

CardsList

  1. H-VLA: Hierarchical Vision-Language-Action Model with Key-Action Reasoning and Motion Planning in a Unified Action Space

    Sep 19, 2026Xiongfeng Peng, Lu Xu, Yandong Wang +10Vision-Language-Action ModelsLanguage-Conditioned Robot Manipulation

    Advanced Research Lab, Samsung R&D Institute China-Beijing (SRCB), China · Samsung AI Center, DS Division, South Korea

  2. TAO-DA: Towards Autonomous Operation--A Dual-Arm Vision-Language-Action Model for Coordinated Manipulation

    Sep 27, 2026Yongsheng Zhao, Han Gao, Baoping Cheng +7Vision-Language-Action ModelsBimanual Robotic Manipulation

    China Mobile (Hangzhou) Information Technology Co., Ltd., Hangzhou, China

  3. DIAL: Decoupling Intent and Action via Latent World Modeling for End-to-End VLA

    Mar 31, 2026Yi Chen, Yuying Ge, Hui Zhou +3Vision-Language-Action ModelsRobotic Manipulation

    The University of Hong Kong · XPENG Robotics · University of North Carolina at Chapel Hill