humanoidsdata.com

Search

Search companies, datasets, articles, and glossary terms for humanoids and embodied AI.

← All glossary terms

Models & learning

Transfer learning

Transfer learning uses knowledge learned from one or more source tasks or domains to improve learning or performance on a different target task or domain. In robotics, transfer may reuse representations, model weights, skills, data, or policies across robots, tasks, environments, sensing configurations, or simulation and reality.

Also known as: knowledge transfer, machine-learning transfer

Updated

Transfer connects a source and a target

Pan and Yang frame transfer learning in terms of source and target domains and tasks. The distinction matters: adapting the same grasping task from simulation to reality is different from reusing a vision representation for a new manipulation task.

Robotics systems transfer pretrained visual features, language representations, dynamics knowledge, motion primitives or complete policies. Fine-tuning is one mechanism, not a synonym for the whole field; transfer can also use frozen features, adapters, shared latent spaces or explicit mappings between embodiments.

Transfer can help or harm

Positive transfer improves the target relative to learning without the source knowledge. Negative transfer occurs when mismatched source assumptions reduce performance. A policy trained on a fixed-base arm may encode reaches or camera geometry that do not carry safely to a mobile humanoid.

Similarity should be stated rather than assumed. Compare action spaces, observations, morphology, controller behaviour, object distributions and success criteria. More source data is not automatically useful if it reinforces the wrong invariances.

Evaluation needs a target-only baseline

Claims of transfer should identify what moved from source to target and what target data remained necessary. Compare against training from scratch or another target-only baseline under the same data and compute budget.

For datasets, retain embodiment and domain metadata so source subsets can be separated. Report target performance across several seeds and conditions, not only the best checkpoint. Transfer is demonstrated by measurable target benefit, not merely by loading pretrained weights.

Sources