RAFT: Reasoning-Aligned Fine-Tuning Improves Robustness of VLA Post-Training
RAFT: Reasoning-Aligned Fine-Tuning Improves Robustness of VLA Post-Training
This research introduces RAFT, a two-stage post-training procedure for vision-language-action (VLA) policies. By first training low-rank adapters on task-relevant reasoning annotations and then freezing them while fine-tuning on action demonstrations, RAFT achieves higher robustness to variations in appearance, objects, and instructions. On a physical DROID robot, it reaches 75.0% out-of-distribution success versus 27.9% for plain fine-tuning. On a multi-stage bimanual YAM task, it attains 70.0% mean success compared to 36.0% for the strongest baseline. The method also demonstrates improved performance under visual, object, and instruction shifts in simulated pick-and-place and open-drawer tasks.
The study presents RAFT, a two-stage post-training method using low-rank adapters on reasoning annotations followed by action demonstration fine-tuning, which improves out-of-distribution robustness and task success on physical DROID and YAM robots compared to plain fine-tuning baselines.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
linear-probe accuracy for object identity
75
View original evidence
otations, then freezes them while fine-tuning the policy on action demonstrations. On a physical DROID robot, RAFT achieves $75.0\%$ out-of-distribution success, compared with $27.9\%$ for plain fine-tuning; on a multi-stage bimanual YAM task, it reaches $70.0\%$ mean success, compared with $36.0\%$ for the strongest baseline. RAFT also improves success under visual, object, and instruction shifts in simulated pick-a
Open source S3
- paper: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
otations, then freezes them while fine-tuning the policy on action demonstrations. On a physical DROID robot, RAFT achieves $75.0\%$ out-of-distribution success, compared with $27.9\%$ for plain fine-tuning; on a multi-stage bimanual YAM task, it reaches $70.0\%$ mean success, compared with $36.0\%$ for the strongest baseline. RAFT also improves success under visual, object, and instruction shifts in simulated pick-a
Open source S3
Source：arXiv Robotics — research abstracts · arxiv.org