LeWAM scaling, arch updates, experiments

Goal: Develop, test, and iterate LeWAM architecture and data pipeline

Work Done:
- create v0.1 and 0.2 arch
- tested both architectures with pretraining, overfitting, policy rollouts, etc.
- set up cloud training infra with vast.ai
- started writing up the paper
- read a ton of WAM and VLA papers for reference
- tested in real life on so101 arm

Successes:

Failures:
- v0.1 was not able to overfit a single frame (forgot to add positional embeddings, plus IDM design was clunky)
- v0.2 was not able to learn simple pick and place task after thousands of steps at large batch size
- burned 200$ on vast.ai :(

Next steps:
- pivoting to new architecture and more ground up approach
- studying for ETH exams so i can resume the project ASAP

Time spent: 100+ hours