Many Ways to Succeed: Diversity-Driven RL Fine-Tuning for VLA Generalization
Many Ways to Succeed: Diversity-Driven RL Fine-Tuning for VLA Generalization
This research proposes DRIVE, a reinforcement learning fine-tuning method for vision-language-action (VLA) systems that leverages successful-behavior diversity to enhance generalization. By grouping rollouts under matched task conditions, aligning trajectories temporally, and deriving an intrinsic reward from relative behavioral diversity, DRIVE improves out-of-domain performance across multiple benchmarks, including LIBERO-Plus, ManiSkill3, and RoboTwin 2.0, with significant gains on a dual-arm AgileX PiPER-X platform.
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
OOD performance
5.3
View original evidence
improves the average out-of-domain (OOD) performance over vanilla RL fine-tuning by 5.3 points on $\pi_0$
Open source S4OOD performance
2
View original evidence
improves the average out-of-domain (OOD) performance over vanilla RL fine-tuning by 5.3 points on $\pi_0$ and 2.0 points on
Open source S4OOD success
9.2
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
$\pi_{0.5}$. On a dual-arm AgileX PiPER-X platform, DRIVE further increases average OOD success from 64.1% to 73.3% (+9.2 points), demonstrating gains that persist under physical deployment.
Open source S5
h temporal alignment, and derives a success-conditioned intrinsic reward from relative behavioral diversity. This design encourages broader coverage of feasible solutions without rewarding diverse failures or superficial timing differences. Across LIBERO-Plus, ManiSkill3, and RoboTwin 2.0, DRIVE improves the average out-of-domain (OOD) performance over vanilla RL fine-tuning by 5.3 points on $\pi_0$ and 2.0 points on
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org