MINERVA: How Small Can a Manipulation Policy Be and Still Solve LIBERO?
The paper introduces MINERVA, a highly compact 0.54M-parameter vision-language-action policy that achieves near-state-of-the-art performance on the LIBERO benchmark, revealing that the task's capacity floor is surprisingly low while exposing critical limitations in instruction conditioning robustness and the lack of benefit from flow matching.
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 are learning to move with a dexterity that once seemed impossible, picking up objects, stacking blocks, and following simple spoken commands. This progress is driven by a new kind of artificial intelligence that combines vision, language, and action into a single system. These systems, often called vision-language-action models, act as the brain of a robot, looking at a scene, understanding a request like "pick up the red block," and then calculating the precise movements needed to complete the task. To train these brains, researchers use a standard set of challenges known as benchmarks, which serve as a measuring stick to see how well a robot performs. For years, the most popular benchmark for robotic manipulation has been a collection of forty different tasks involving moving objects in a simulated environment. The prevailing belief in the field has been that to solve these tasks well, a robot needs a massive brain—a digital model with billions of parameters, or internal settings, that requires powerful, expensive computer hardware to run.
However, a team of researchers from the University of Tokyo has asked a simpler, more practical question: just how big does the brain actually need to be? If a robot is deployed to do a specific set of jobs in a factory or a home, it does not need the ability to understand every language in the world or recognize every object it has never seen before. It only needs to solve the specific tasks it was hired to do. The researchers wanted to find the smallest possible version of a robot brain that could still solve the standard set of forty tasks perfectly. They were looking for the minimum amount of computing power required to get the job done, a threshold that would determine whether a robot could run on a small, cheap chip or if it would always need a massive server.
To find this limit, the team built a family of robot policies, which are the programs that tell the robot how to move, and systematically made them smaller. They started with a model containing nearly ten million internal settings and began to shrink it, step by step, watching to see when the robot would start to fail. They stripped away complex features that are common in larger models, such as the ability to read natural language sentences or use pre-trained vision systems. Instead, they gave the robot a simple list of forty task names, represented by numbers, and a camera that learned to see from scratch. They then trained these models on the standard forty tasks and tested them over two thousand times to see how often they succeeded.
The results revealed a surprising floor. The researchers found that a model with just 0.54 million parameters was able to solve the tasks with a 95.1 percent success rate. This tiny model performed almost as well as the largest, most famous models in the field, which contain billions of parameters and are thousands of times larger. The performance of the robot did not improve significantly once the model size reached about one million parameters; adding more computing power beyond that point yielded diminishing returns. Conversely, when the model was shrunk below a quarter of a million parameters, the robot's ability to solve the tasks collapsed, particularly on the more complex, long-duration tasks. This suggests that for this specific set of challenges, the robot does not need a giant brain; it only needs a small, efficient one.
The study also uncovered what parts of the robot's brain actually matter. The researchers tested different ways of organizing the model and found that the most critical resource was the part that processes visual information—the "eyes" of the robot. If they took away too much capacity from the visual system, the robot failed, regardless of how much power was left for the part that planned the movements. They also discovered that the complex mathematical methods often used to generate smooth, fluid movements were not necessary for this level of performance. A simpler, faster method of calculating movements worked just as well and allowed the robot to make decisions in a fraction of a second.
Perhaps the most striking finding was how the robot understood instructions. In the larger models, the robot reads a sentence like "pick up the cup" and tries to understand the meaning of the words. In this tiny model, the researchers replaced the language with a simple number code. When they scrambled these codes so that the robot was given the wrong number for a task, the robot's success rate dropped to near zero. This proved that on this specific benchmark, the robot was not truly "understanding" language in a general sense; it was simply memorizing which visual pattern matched which number code. The instruction was just a key to unlock a specific memorized behavior.
This distinction has profound implications for how robots are built and used. The researchers showed that the high success rates reported by giant models on this benchmark are largely a measure of how well the robot has memorized the specific tasks, rather than how well it can generalize to new situations. When they tested these tiny models against variations of the tasks—such as changing the lighting, the background, or the shape of the objects—the success rate dropped dramatically, revealing that the models had not learned the underlying physics of the world, only the specific look of the training tasks. However, for a robot that is deployed to do the same forty tasks every day, this memorization is exactly what is needed.
The practical outcome of this research is a robot policy that is incredibly efficient. The 0.54 million parameter model can run on a standard laptop computer without any specialized graphics cards, making decisions in less than ten milliseconds. This is hundreds of times faster than the current state-of-the-art models, which often take seconds to plan a single move. By proving that a tiny, specialized model can solve the standard benchmark, the researchers have shown that the path to practical, affordable robots does not necessarily require building ever-larger brains. Instead, it requires finding the smallest, most efficient brain that fits the specific job, a discovery that could allow robots to be deployed in homes and factories where space, power, and cost are limited. The study does not claim that these small models can replace the large, general-purpose systems needed for open-ended exploration, but it firmly establishes that for a fixed set of tasks, the capacity required is far smaller than previously believed.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.