Joint Team Including SoftBank Corp. Members Wins Robot Foundational Model Development Competition

Imagine a future where robots understand human instructions and perform everyday tasks autonomously. Research and development in the field of Physical AI, which combines robotics and AI, is bringing this vision closer to reality.

Working to help realize this vision, a team of Physical AI researchers-comprising members from SoftBank Corp. (TOKYO: 9434), the National Institute of Advanced Industrial Science and Technology (AIST) and Mitsubishi Electric Corporation-took first place in the Robot Foundational Model Development Competition, an event organized by the AI Robot Association (AIRoA) that ran from April 15, 2025, to February 16, 2026.

SoftBank News spoke with SoftBank participants in the joint team to learn more about the development efforts and innovations that led to their victory, and how they helped robots perform everyday tasks with precision.

Using massive volumes of training data to realize precise robot performance

What is the Robot Foundational Model Development Competition?

Hori: It's a competition that challenges participants to develop AI models for service robots that assist people in their daily lives.

Generative AI capable of processing text and images is already well known, but this competition focused on the next frontier-enabling AI to operate in the real world. To do this, robots use information gathered from cameras and other sensors to understand their surroundings, determine the appropriate actions and execute them. Each team developed the AI model that serves as the robot's "brain."

Using tens of thousands of hours of robot operation data collected by AIRoA, the teams competed to see how successfully their robots could perform everyday tasks. Rather than measuring performance with simulations alone, the competition evaluated success rates of tasks performed in the real world, making the challenge highly practical.

A robot recognizes and grabs a handkerchief in a container

What kinds of everyday tasks were the robots expected to perform?

Suzuki: Tasks included things like grasping and lifting objects and opening a microwave door. These actions may be effortless for people, but they're extremely difficult for robots. Food items vary in shape and firmness, while clothing is soft and never exactly the same shape twice. A robot must determine where to look for an object, where to grasp it, and how much force to apply. Every movement requires the robot to process and make decisions based on a wide range of data.

What role did SoftBank play as part of the joint team?

Hori: SoftBank was primarily responsible for AI model training and evaluation, data processing, building the training pipeline and validating performance through simulation.

Robot operational data spans a wide variety of information, including video, joint movements, object positions, and whether each task was successfully completed. We processed this data into a format suitable for AI training, verified robot behavior in a simulation initially, and then tested the model on physical robots once the simulation results were satisfactory.

When people think about robotics, they often focus on the hardware itself. Behind the scenes, however, developing robots requires systems capable of processing enormous volumes of data, continuously training AI models and evaluating their performance. SoftBank contributed its expertise in AI, data processing, and computing infrastructure to support the foundational technologies that underpin AI development.

What was the biggest challenge during the competition?

Nishihara: Our greatest challenge was repeatedly cycling through data processing, model training, evaluation and improvements within a limited timeframe.

The dataset provided for the competition consisted of unprocessed, "raw" data. The volumes were so large that training on the entire dataset would have taken more than a month, and the quality of the data varied considerably. We had to train the models, test them on robots, identify which changes actually improved performance, and then decide what to modify next based on the results. Repeating this validation process many times in a short period made the project both technically demanding and time-consuming.

Suzuki: The challenge was more than just dealing with the sheer volume of data. Another major hurdle was translating the trained model's performance into successful operations on an actual robot.

With robots, good results in a simulation don't necessarily translate into success in the real world. Physical errors can also occur-for example, even if you instruct a robot to move one centimeter, it may move only nine millimeters. There were times when we spent one to two weeks training a model, only to find that the training failed when deployed on a physical robot.

Identifying the data that matters most

What do you think was the key factor behind your success in the competition?

Suzuki: One of the biggest factors was that we didn't simply use the enormous volume of robot operational data as it was. Instead, we identified which data would be most effective for training the model.

In AI development, it's often assumed that more data automatically leads to better performance. However, that's not necessarily the case when it comes to robot operation data-not every data point contributes equally to improving a model. So we focused on evaluating the quality of the data and just selected the samples that were most valuable for training. Our goal was to build a higher quality model capable of achieving a greater success rate for performing tasks in real-world environments.

To identify high-quality data, we first needed a way to assess what "high quality" actually meant. We used AI to compare images captured when a task was successfully completed with what the robot's camera was seeing during execution. Based on how direct and efficient the robot's movements were, we assigned each sequence a score, treating higher-scoring data as better suited for training.

Nishihara: After scoring the data, we also paid close attention to how we selected the training dataset. For example, if we trained the model using only perfect, top-scoring examples, even a slight positional error in the actual robot could cause the AI to fail because it hadn't learned how to handle deviations.

To improve resilience against the inevitable inaccuracies of physical robots, we intentionally mixed in some lower-scoring data. This careful balance allowed the robot to tolerate small variations and perform more reliably in real-world conditions.

How are these results significant?

Hori: We believe our results demonstrate just how important data quality and data selection are for improving robot foundational models.

While designing better AI model architectures is certainly important, for AI systems that operate in the physical world-like robots-the choice of training data has a tremendous impact on real-world performance. Through this competition, we showed that carefully curating training data can lead to higher task success rates when robots perform actual physical tasks.

Suzuki: We also believe this work has provided valuable insights for advancing Physical AI.

Traditionally, robots have been designed to repeat the same tasks in controlled environments like factories. As they become capable of more general-purpose behavior, however, they have the potential to assist people in homes, retail stores and many other everyday settings.

In this competition, the challenge was to develop AI models capable of making those kinds of general-purpose decisions. Rather than being evaluated solely through simulations or offline benchmarks, the models were judged on whether they could successfully complete tasks in the real world. The experience gave us confidence in how data should be selected and how training should be approached to develop more capable Physical AI systems.

(Posted on October 1, 2026)
by SoftBank News Editors