NVIDIA VFlash speeds up robot trajectory decoding by 12x

Breaking the sequential bottleneck

Current vision-language models (VLMs) often struggle with speed when they need to map out complex spatial movements. These models typically generate coordinates one at a time, a process known as autoregressive decoding. This sequential approach becomes a major bottleneck as robots are asked to track more objects or plan longer, more detailed paths.

NVIDIA researchers addressed this limitation by developing VFlash. This new visual planning method allows models to generate complete trajectories and other spatial predictions in a single pass. By moving away from the coordinate-by-coordinate method, the system can handle increasingly detailed spatial information without a massive spike in processing time.

The primary goal of VFlash is to improve how robots interpret and interact with the physical world. As robots move from controlled environments to unpredictable human spaces, they need to process visual data quickly. VFlash aims to provide the speed necessary for real-time interaction while maintaining the high level of accuracy required for safe movement.

How Flash Primitive Decoding works

VFlash operates by identifying the specific structure of a requested visual prediction before it actually generates the coordinates. For instance, the model recognizes that a bounding box has a fixed coordinate layout, while a trajectory might contain a variable number of points in two or three dimensions. Once the structure is identified, the model knows exactly what kind of data it needs to produce.

To generate these coordinates, VFlash uses a lightweight diffusion-based drafter. This component allows the model to predict the necessary points jointly in one forward pass. This is different from conventional speculative decoding, which requires the system to check every generated token against a target model to ensure they match.

Because VFlash understands the underlying geometry, it can accept structurally valid visual predictions without needing token-by-token agreement. This architectural shift allows for much faster output. In preliminary single-rollout measurements, the researchers found that VFlash could generate predictions approximately 25 times faster than standard autoregressive decoding in certain tasks.

Performance gains in 3D space

The researchers tested VFlash across various two-dimensional and three-dimensional robot trajectory prediction tasks. The results showed a significant leap in accuracy. In three-dimensional trajectory tests, the method lowered the average displacement error by as much as 61%, depending on the specific dataset and how the model was trained.

The speed improvements were also substantial during robot manipulation tests. On NVIDIA’s Jetson Thor platform, the system's processing time dropped from 3,552.7 milliseconds to 703.3 milliseconds. This represents an approximately 80% reduction in the time it takes for the model to process the information compared to the baseline system.

Similar results were seen on DGX Spark hardware, where latency dropped from 3,356.5 milliseconds to 649.9 milliseconds. These numbers suggest that robots equipped with this technology can respond to changing environments much more rapidly. This speed is critical for tasks where a delay of even a few milliseconds could result in a failed movement or a collision.

Impact on robotic success rates

Beyond just speed and accuracy, the researchers wanted to see if faster predictions actually led to better physical performance. They linked the trajectories produced by VFlash to a world action model that was tested using the MolmoSpaces benchmark for robotic manipulation. The results showed that the model achieved an average success rate of 70.2% across four different task suites.

This was a notable improvement over the baseline system, which had a success rate of 62.6%. The most significant gains were observed in specific types of movements. For pick-and-place tasks, the success rate increased by 11.6 percentage points, while object-closing tasks saw an increase of 12.8 percentage points.

These findings indicate that the efficiency of VFlash translates directly into better physical outcomes for robots. By providing more accurate and faster plans, the system helps the robot complete complex maneuvers more reliably. This suggests that the method is effective for practical applications where robots must interact with objects in a sequence of precise actions.

Applications in autonomous driving

While the primary focus of the research was on robotics, the researchers also explored how VFlash could be applied to autonomous driving. They tested the model using the NAVSIM and Waymo datasets, which are standard benchmarks for evaluating self-driving technology. The model was able to generate driving trajectories that performed competitively against existing benchmarks.

To further refine the results, the researchers utilized reinforcement learning to improve the planning outcomes. This process helps the model learn from experience to make better decisions in complex traffic scenarios. The findings suggest that the ability to generate detailed visual plans efficiently is a versatile tool for various types of physical AI.

As interestingengineering.com reported, the research team published their findings in a paper titled VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding in November 2026. It highlights a move toward systems that can handle high-dimensional spatial data in real-time. As these models become more integrated into hardware, they could help robots and vehicles navigate the world with greater fluidity and precision.