A small policy module raised a pretrained robot model’s average success from 37.5% to 71.3% across four laboratory manipulation tasks by changing the noise fed into its action generator, rather than retraining that generator, a new arXiv preprint reports. The tasks were cup stacking, toy fishing, sequential insertion and bead placement, with 40 evaluation attempts per task.
The method, called noise space learning, works with vision-language-action models that turn camera images, robot state and an instruction into movements. In flow-matching versions, a generator starts from random noise and gradually turns it into a sequence of actions. The researchers kept that generator fixed and trained a smaller, context-aware policy to change how the starting noise was sampled, steering the robot toward different movements the pretrained model could already produce.
In autonomous tests, the adapted policy averaged 71.3% success across the four tasks, compared with 37.5% for the initial policy, 58.8% for Action-AWR and 70.0% for residual adaptation. When researchers supplied human corrections, average success reached 83.8%; the paper reports 66.3% for DAgger and 73.8% for UniSteer in that human-guided comparison. The averages weight each task equally, so they describe this small set of lab trials rather than a general deployment rate.
The paper also reports low model-update time: 1.53 minutes per task on average for autonomous adaptation, versus 52.4 minutes for Action-AWR and 23.2 minutes for residual adaptation. That measure covers cumulative time spent updating the model, not robot data collection or evaluation, so it is not the total time needed to adapt a robot.
Keeping the main generator fixed did not make every update safe. The authors report that an earlier real-robot run of their SAC variant caused unintended contact with the table during bead placement, so they used the AWR variant for their physical comparisons. They also warn that noise distributions far from the usual Gaussian starting point can produce unsafe motions and trigger safety interventions.
The results come from four lab tasks in a single arXiv v1 preprint. They show that a small learned module can reshape a frozen model’s action distribution in these tests; they do not establish reliable performance across other robots or guarantee safe behavior.
Sources
- Junjie Lu et al., “Noise Space Learning for VLA Adaptation,” arXiv:2610.13943v1, submitted Oct. 9, 2026.
- Illustrative photo, not from the study: robot arm assembling a component in a lab by Nenad Stojković, licensed under CC BY 2.0.
