An October 7 arXiv preprint introduces DRIVE, a reinforcement-learning method for fine-tuning vision-language-action (VLA) robot controllers, systems that turn camera views and language instructions into movements. In tests on a physical dual-arm robot, the authors report higher success when the scene changes than with standard fine-tuning. Each task-condition pair was tested 20 times.
Standard task rewards say whether a robot succeeded, but not whether it used a different successful approach. DRIVE groups attempts from the same task and matched scene conditions, then compares sequences of features the VLA policy computes from its visual and language inputs at each decision. Temporal alignment accounts for differences in execution speed, so a slower attempt is not automatically counted as a new behavior. Only successful attempts receive a diversity bonus.
On hardware, two AgileX PiPER-X arms performed Click Bell and Press Stapler. The team first trained both methods from the same supervised-fine-tuning starting point in a simulation aligned to the physical setup. They tested each task in a clean scene, then with distractor objects, a changed tabletop background or changed lighting; each method had 20 trials for every task-condition pair.
Across the six changed-scene combinations (three scene changes across two tasks), DRIVE averaged 73.3% success; vanilla reinforcement-learning fine-tuning averaged 64.1% across those same combinations. Including clean scenes, DRIVE scored higher in seven of eight task-condition comparisons and tied in one.
The physical evidence covers two tasks on one robot setup. The authors say pairwise trajectory comparisons add computation and the method depends on the VLA model's visual-language features. They call for more efficient metrics that better distinguish task-solving strategies from incidental movement, and further tests on longer tasks and more robot designs.
Sources
- Li et al., "Many Ways to Succeed," arXiv:2610.09943v1 (October 7, 2026).
- Lead image: Figure 6 from the authors' paper, reused under CC BY 4.0.