Estonian researchers teach underwater robots to learn from bigger kin
No labeled data needed—AI transfers dynamics from large to small underwater robots.
Researchers from TalTech (Estonia) have proposed an efficient transfer learning framework for modeling the dynamics of soft, fin-actuated underwater robots. Led by Pavlo Kupyn, the team focused on morphologically similar robots that differ in scale and hydrodynamic properties. Their approach uses an autoencoder-based domain adaptation method to learn a shared latent representation that aligns the dynamics of a larger source robot with a smaller target robot. This allows a model trained on the larger robot's data to be adapted to the smaller one using limited or even zero labeled data.
Experiments on two real underwater robots demonstrated that the method achieves accurate body-frame velocity state estimation on the target platform without any labeled data. The work was accepted for publication at the 2026 12th International Conference on Control, Decision and Information Technologies (CoDIT). This technique could significantly reduce the data collection burden for deploying new robot variants, especially in environments where obtaining labeled training data is expensive or impractical.
- Autoencoder-based domain adaptation aligns dynamics of differently scaled underwater robots.
- Achieves zero-shot state estimation on smaller target robot without labeled data.
- Accepted at CoDIT 2026; tested on two real soft fin-actuated platforms.
Why It Matters
Slash data collection costs for underwater robot deployment by reusing models across similar platforms.