Lifelong robot learning is challenging because robots must learn new tasks sequentially without losing the ability to perform earlier ones, a problem often called catastrophic forgetting. The paper introduces SARAM, which stands for Sparse Adapter Allocation with Reuse and Merging, as a way to address this issue.
The method builds on large pre-trained vision-language-action (VLA) backbones. Instead of fine-tuning the entire model, SARAM uses adapter-based techniques, adding small trainable modules that can be allocated, reused, and merged across tasks. This sparse approach is designed to mitigate forgetting while keeping the process efficient.
The abstract indicates that adapter-based methods effectively handle this challenge, but it does not provide specific experimental results or comparisons. The full paper is available on arXiv for those seeking further details.