IWAI 2023 - Tetsuya Ogata - Embodied AI with Active Inference
Tetsuya Ogata, a robotics researcher at a university in Japan, discusses building robots that adapt through prediction and active inference, with examples from his research and collaborations in Japan.
In this talk
- 2:26
Why physical robotics remains difficult
Ogata contrasts deep learning’s strong results in areas such as images and language with the difficulty of applying it to real hardware. He says common approaches like object recognition and reinforcement learning need more training data than is practical to collect for many robot tasks.
- 5:25
Learning motions by reducing prediction error
Ogata describes a robot that learns towel folding using 50 examples, improving a task that initially takes over a minute to about 10 seconds. Its neural networks predict future images and motion, then the robot changes its movement when the real scene differs from its prediction. The system also handles unfamiliar objects, though strong expectations can lead it to try folding a desk when towels are absent.
- 17:38
Applying predictive control to varied tasks
Examples include opening a jar with 36 teaching examples, handling materials with changing properties, and opening doors while maintaining balance. In a food handling project with Japanese industrial companies, the robot explores material by touch before choosing a tool. Ogata says combining visual, force, and tactile signals helps it adapt its actions.
- 28:04
Adding uncertainty and scaling to general-purpose robots
Ogata says his current models emphasize reducing prediction error, while future work should also address uncertainty. A free-energy model can adjust internal expectations during motion, but he notes that it carries high computational costs. In Japan’s Moonshot Goal 3 project, his team is working toward robots that can clean, cook, and wash; an egg-cooking robot shifts attention from vision toward touch and torque as the egg changes.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.