💻 computer science

DriftingVLA: Native One-Step Vision-Language-Action Generation via Per-Dimension Temporal Drifting

DriftingVLA is a native one-step vision-language-action model that utilizes a distribution-drifting objective with Per-Dimension Temporal Drifting to generate complete action chunks in a single forward pass, achieving state-of-the-art performance on benchmark tasks while delivering a 3.36-fold speedup over conventional multi-step flow-based methods.

Original authors: Yuxuan Gao, Shiqi Zhang, Yedong Shen, Yifan Duan, Wenhao Yu, Xin Zhang, Siyuan Cao, Jiajun Deng, Yanyong Zhang

Published 2026-09-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuxuan Gao, Shiqi Zhang, Yedong Shen, Yifan Duan, Wenhao Yu, Xin Zhang, Siyuan Cao, Jiajun Deng, Yanyong Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Robots that can see, understand language, and move with human-like fluidity have long been the dream of artificial intelligence. For years, researchers have built systems that combine powerful visual and linguistic understanding with the ability to control robot arms. These systems, known as vision-language-action models, act as the brain of the robot, taking in what the camera sees and the words a human speaks, then deciding how to move the robot's joints to complete a task. However, a significant bottleneck has held these machines back from true real-time responsiveness. To generate a smooth sequence of movements, these models traditionally rely on a complex, multi-step process. Imagine trying to draw a perfect circle by making tiny, hesitant corrections at every single point along the line; the robot has to calculate a path, take a step, check its work, and then take another step, repeating this cycle many times for every single movement it makes. While this method produces high-quality results, it is slow, introducing a delay that makes the robot feel sluggish and unresponsive in a dynamic world.

A team of researchers has now introduced a new approach that fundamentally changes how these robots plan their movements, allowing them to generate a complete sequence of actions in a single instant. Their work, centered on a system they call DriftingVLA, replaces the slow, iterative correction process with a method that learns to predict the entire future movement at once. Instead of calculating a path step-by-step during the actual task, the system learns a direct connection between a random starting point and the final desired motion during its training phase. This shift allows the robot to skip the repetitive calculations entirely when it is deployed, jumping straight to the correct action. In tests involving complex manipulation tasks, this new method not only matched the precision of the older, slower systems but also made the robot more than three times faster, reducing the time it takes to decide on a movement from over two hundred milliseconds to under seventy.

The core of this breakthrough lies in how the researchers taught the robot to understand the relationship between its different moving parts. A robot arm does not just move in a single line; it has multiple joints and degrees of freedom that must work together, such as moving forward, rotating, and opening a gripper. Previous attempts to speed up these systems often treated the entire movement as one big block or broke it down into tiny, disconnected moments in time. The new approach, however, recognized that each specific direction of movement—like the vertical lift of a gripper or the rotation of a wrist—has its own unique rhythm and statistical pattern. The researchers developed a technique called Per-Dimension Temporal Drifting, which treats the full history of each individual movement direction as a separate unit to be learned. This allows the system to understand the specific nuances of how a gripper should open or how a wrist should turn without forcing these different motions to conform to a single, rigid mathematical rule.

Crucially, this method does not sacrifice the robot's ability to coordinate its limbs. Even though the system learns the patterns of each movement direction separately, it still generates the entire action plan as a unified whole. The robot's brain, which understands language and vision, remains intact and continues to guide the action expert that produces the movements. By training the system to refine its predictions across many different "what-if" scenarios simultaneously, the researchers ensured that the final, single-step decision is just as accurate as the result of a slow, multi-step calculation. This means the robot can still perform delicate tasks, like pouring liquid from one container to another or stacking blocks with two arms working in sync, without the hesitation that plagued earlier models.

The researchers tested this new system in both simulated environments and the real world to see if the speed came at the cost of reliability. In a series of challenging simulations involving tasks like picking up objects and rearranging them, the new system achieved a success rate of nearly 98.3 percent, slightly outperforming the best existing multi-step models. When moved to a real-world setting with physical robot arms, the system maintained its edge, succeeding in 77.67 percent of trials across six different tasks, including both single-arm and dual-arm coordination challenges. This performance was notably higher than other one-step methods that tried to speed up the old systems by simply cutting short their calculation process, which often resulted in a significant drop in accuracy. The new method proved that by changing how the system learns, rather than just how it calculates, it is possible to achieve both speed and precision.

Beyond the raw numbers, the study highlighted the efficiency gains that come from removing the repetitive calculation loop. In the real world, where every fraction of a second counts, the new system reduced the time required to generate a movement chunk from 227.61 milliseconds to 67.67 milliseconds. This represents a speedup of more than three times, effectively eliminating the lag that makes robots feel disconnected from their environment. While the training process required slightly more computer power to handle the multiple "what-if" scenarios during the learning phase, this cost is a one-time investment. Once the robot is trained, it operates with a lean, single-step efficiency that does not demand extra resources. The work demonstrates that the future of responsive robotics may not lie in building faster computers to run slower algorithms, but in redesigning the algorithms themselves to be inherently direct and immediate.

Technical Summary: DriftingVLA

Problem Statement

Conventional Vision-Language-Action (VLA) models, particularly those based on flow matching (e.g., $\pi_0$ , $\pi_{0.5}$ ), excel at generating expressive, continuous action chunks. However, they rely on iterative numerical integration at inference time to recover actions from a learned vector field. This process requires multiple sequential evaluations of the action expert (e.g., 10 Network Function Evaluations or NFEs), introducing significant latency that hinders real-time online robot control.

While recent efforts have attempted to accelerate these policies via distillation (e.g., SnapFlow) or by adapting flow formulations for single-step prediction (e.g., MeanFlowVLA), these methods often remain derived from the inference-time transport process. There is a need for a native one-step VLA that generates a complete action chunk in a single forward pass without relying on iterative refinement or numerical solvers during deployment, while maintaining the high control performance of multi-step flow policies.

Methodology

The authors propose DriftingVLA, a framework that shifts the burden of distribution refinement from inference-time integration to training-time optimization using Distribution Drifting.

1. Native One-Step Generation via Distribution Drifting

Unlike flow-based models that learn a velocity field $v_\theta$ to be integrated over time, DriftingVLA learns a direct mapping $f_\theta$ from Gaussian noise to a complete action chunk.

Training Objective: The model utilizes a Drift-Based Policy (DBP) formulation. During training, for a given context $c_t$ , the model generates $G$ sibling action chunks from independent latent samples. A "drifting field" is constructed using attraction toward demonstrated trajectories and repulsion from generated siblings.
Optimization: Instead of applying the drifting field iteratively at inference, the corrections are absorbed into the generator parameters via a stop-gradient regression objective. Once trained, the policy executes a single forward pass: $\hat{A}_t = f_\theta(Z_t, c_t)$ , where $Z_t$ is a single noise sample.
Architecture Adaptation: DriftingVLA retains the pretrained multimodal backbone (VLM) of $\pi_{0.5}$ but freshly initializes the action expert. The action expert is retrained to map noise directly to action chunks rather than predicting flow velocities. The VLM and action expert are jointly post-trained, with the multimodal context shared across sibling samples to avoid redundant computation.

2. Per-Dimension Temporal Drifting (PDTD)

A critical challenge in extending distribution drifting to VLAs is defining the geometry of the action chunk $A \in \mathbb{R}^{H \times D}$ (where $H$ is the temporal horizon and $D$ is the action dimension).

Limitations of Existing Organisms:
- Chunk-wise Drifting: Flattens the entire chunk, mixing heterogeneous action channels (e.g., translation, rotation, gripper) into a single geometry. This causes cross-channel interference where the scale of one dimension affects the drifting signal of others.
- Step-wise Drifting: Treats each time step independently, fragmenting the temporal trajectory of individual action channels.
PDTD Solution: The authors introduce Per-Dimension Temporal Drifting, which treats the complete temporal trajectory of each individual action dimension ( $A_{:,d} \in \mathbb{R}^H$ $A_{:, d} \in R^{H}$ ) as a separate drifting unit.
- Geometry: Drifting statistics (distance scales, affinities, forces) are computed independently for each action dimension $d$ .
- Joint Generation: Crucially, PDTD factorizes only the training geometry, not the generator itself. The shared VLA model still generates the full $H \times D$ chunk jointly, preserving cross-dimensional dependencies required for coordinated control.

Key Contributions

Native One-Step VLA: The extension of native drift-based generation to pretrained flow-based VLAs. This is achieved by preserving the pretrained multimodal backbone while relearning the action interface as a direct conditional generator, eliminating the need for numerical integration or flow-trajectory distillation at deployment.
Per-Dimension Temporal Drifting (PDTD): A novel organization for action-chunk drifting that respects the distinct control semantics and distributional characteristics of different action dimensions. PDTD preserves complete per-channel temporal trajectories while separating heterogeneous channel geometries, outperforming both Chunk-wise and Step-wise approaches.
Comprehensive Evaluation: An extensive evaluation across simulation benchmarks (LIBERO, RoboTwin 2.0) and six real-world single- and dual-arm manipulation tasks, demonstrating that native one-step generation can match or exceed the performance of iterative flow policies while significantly reducing latency.

Experimental Results

Simulation Performance:
- LIBERO: DriftingVLA achieved a 98.32% overall success rate, outperforming the original 10-NFE $\pi_{0.5}$ policy (97.10%) and the one-step baseline SnapFlow (97.48%).
- RoboTwin 2.0: DriftingVLA reached 81.09%, exceeding 10-NFE $\pi_{0.5}$ by 1.51 points and SnapFlow by 2.56 points. Notably, a standard one-step truncation of the standard $\pi_{0.5}$ policy dropped to 63.76%, highlighting the necessity of native training.
Real-World Performance:
- Across six real-world tasks (single and dual-arm), DriftingVLA achieved a 77.67% success rate, outperforming $\pi_{0.5}$ (74.22%) and other one-step baselines by significant margins.
- PDTD consistently outperformed Chunk-wise and Step-wise variants in both simulation and real-world settings.
Efficiency:
- Latency: Native one-step generation reduced action-chunk latency from 227.61 ms (10-NFE) to 67.67 ms, a 3.36× speedup.
- Training Overhead: Increasing the number of sibling samples ( $G$ ) from 2 to 8 increased training GPU-hours by only 3.78% and peak memory by 9.54%, due to the shared-context sampling design.

Significance

The paper demonstrates that iterative action refinement can be removed from the deployment loop without sacrificing control performance. By moving distribution refinement to the training phase via distribution drifting and utilizing a per-dimension temporal geometry, DriftingVLA achieves native one-step deployment. This approach resolves the latency bottleneck of flow-based VLAs while maintaining the expressiveness and coordination capabilities required for complex robotic manipulation, offering a viable path toward real-time, high-performance robot control.