SmolVLA with and without a LoRA adapter

These are pre-rendered rollouts. The model does not run in your browser. Every clip is a real simulator episode recorded during the evaluation (same seed and same object layout for both policies), chosen by episode index, not by outcome; failures are included. The result of the study is a null result: see the table below and the README.

Instruction:

Success per task (all recorded evaluation episodes)

Task 5 is the target (90 episodes per policy: 3 evaluation seeds x 30). Other tasks are a small forgetting check (10 episodes each); too few to show small changes.