Compositional and Modular Learning in the Era of Scaling in Robotics

Compositional or modular learning enables systems to exploit the structure of the problem by building abstractions between functionally different components. These structural assumptions of the underlying problem help in reducing the sample complexity of the learning system.

This has led to increased adoption of compositional approaches within the robotics community, primarily due to the diversity and the lack of standardization in robot embodiments, observational modalities and task goals that make collection of data supporting all combinations difficult.

As various efforts attempt to scale data and build large foundation models, it could be argued that these foundation models approximate compositional behavior for the scaled modality, embodiment and goal combinations. However, certain modalities such as force and tactile are intrinsically difficult to scale and standardize across different setups. Moreover, human actions are highly compositional, dextrous and goal oriented, making the scaling of the training data along the skill axis a costly endeavour.

We propose this workshop to solicit novel works that use the principle of compositionality or modularity for tackling learning problems in robotics. Specifically, we encourage works that leverage the representational power of scaled systems in their approach to showcase the best of both worlds.