Brain theory-based AI enables robot to learn multiple caregiving tasks

A research group including Specially Appointed Research Unit Leader Hayato Idei and Unit Leader Yuichi Yamashita at the Department of Information Medicine, National Institute of Neuroscience, National Center of Neurology and Psychiatry (NCNP), alongside Junior Researcher Tamon Miyake and Professor Tetsuya Ogata of Waseda University, has implemented an AI model based on brain theory in a multi-degree-of-freedom humanoid robot. The researchers demonstrated that a humanoid robot can learn multiple caregiving tasks while integrating high-dimensional visuo-proprioceptive sensory information. The findings were published in Science Advances.

Although AI is advancing rapidly, it remains difficult for AI systems to integrate vast amounts of information, such as visual and bodily sensory inputs, and flexibly switch between different tasks according to the situation, as humans do.

One computational principle attracting attention as a possible basis for such abilities is predictive processing, also known as the free-energy principle. The research group tested this brain theory experimentally.

Under this theory, the brain does not passively process information from the external world. Instead, it actively predicts future sensory inputs and performs perception, action, and learning by minimizing the discrepancy between those predictions and actual sensory inputs, known as prediction errors. However, whether this theory can function effectively in high-dimensional, multisensory, and multitask environments such as real-world robots has not been sufficiently verified.

To examine whether predictive processing remains effective in such environments, the researchers focused on robot learning of caregiving tasks. Labor shortages in caregiving settings are becoming increasingly severe and represent an urgent societal challenge. In caregiving tasks such as repositioning patients and body wiping, actions must be flexibly adjusted according to changes in both the care recipient and the surrounding environment while integrating visual and bodily sensory information. Generating such behavior in caregiving scenarios is also one of the most challenging tasks for evaluating adaptive intelligence in real-world environments.

The researchers therefore developed a new predictive processing-based AI model, Scalable PV-RNN, and implemented it in the multi-degree-of-freedom humanoid robot AIREC to evaluate whether the theory could support high-dimensional multitask intelligence. The AI directly integrates approximately 30,000 dimensions of visual and proprioceptive information and learns while predicting future sensory inputs.

As a result, the researchers showed that complex multisensory information can be integrated using a single computational principle, minimizing prediction errors, without relying on feature extraction, dimensionality reduction, attention mechanisms, or other techniques commonly used in conventional robotic AI systems.

In the experiments, the researchers used data obtained by teleoperating AIREC and trained the robot on two fundamentally different caregiving tasks: repositioning a mannequin care recipient and body wiping.

The results confirmed that the proposed AI model can simultaneously learn and predict these different caregiving tasks. Unlike conventional approaches that require separate AI systems for each task, the findings suggest that a single brain-inspired AI system could adapt to a variety of different tasks.

Furthermore, analysis of the internal structure of the AI model revealed that information-processing characteristics known to occur in the human brain had emerged within the model. These included: (1) estimating situations from bodily sensations even when visual information is ambiguous; (2) autonomously learning when to switch between actions; (3) inferring visually occluded objects; and (4) recognizing changes in uncertainty depending on the situation.

These functions were not explicitly designed into the system but emerged naturally through learning based on predictive processing, suggesting that the theory could provide a computational principle for supporting real-world, high-dimensional multisensory intelligence. The researchers expect the approach to serve as a computational foundation for future robotic control systems.

Beyond advancing robotic AI, the findings also provide new insights into a fundamental question in neuroscience: how the human brain achieves flexible intelligence through its underlying computational principles.

Journal Information
Publication: Science Advances
Title: Predictive processing as a scalable computational principle for embodied multitask intelligence
DOI: 10.1126/sciadv.aed7511