Reproduce “RAFT: Reasoning-Aligned Fine-Tuning Improves Robustness of VLA Post-Training” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shresth Grover, Andrey Kolobov, Dhruv Shah
RAFT adapts a vision-language-action policy to robot tasks through two stages. It first trains low-rank adapters to predict task-relevant and spatial information, then freezes them while training the action policy. In DROID pick-and-place, RAFT achieves 75.0% success across 68 trials pooled over three out-of-distribution shifts, versus 27.9% for fine-tuning. This supports robustness in this setting, not a universal guarantee.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.