Catastrophic forgetting

New training can interfere with old skills#

Overcoming Catastrophic Forgetting in Neural Networks studies the difficulty of learning tasks sequentially while retaining earlier expertise. Shared parameters that supported an old task can change when optimization focuses on a new one.

For a robot-policy example, fine-tuning on a new insertion task could reduce performance on an earlier grasping task. That is a possible form of forgetting; it must be measured rather than inferred simply because the model was updated.

Preserve parameters important to earlier tasks#

The paper introduces elastic weight consolidation, which slows changes to weights estimated to be important for previously learned tasks. Its experiments include sequential image-classification tasks and Atari games.

This is one proposed mitigation. The reported results do not establish that every robot skill can be retained while adding unlimited new capabilities.

Evaluate earlier tasks after each update#

A new-task success score alone cannot reveal forgetting. Assessment needs comparable evaluations of both the new task and the old tasks after adaptation.

For a robot foundation model, this distinction matters when a broadly trained model is specialized for one platform or environment. Improvement on the new setting and retention of previous abilities are separate outcomes. Changes in hardware or evaluation conditions should also be separated from losses caused by model training.