Training and Inference
Here, adapted backbone features, the robot state, and the robot embodiment are used to condition action prediction. The loss averages squared error over valid action dimensions and timesteps. State and action values are normalized using the 1st and 99th percentiles of the training data. Gradients update only the action head, so backbone features are precomputed once and reused during training.
Experiments
Policies with vision-language model backbones ranging from 2B to 552B parameters are compared. The experiments measure task success in simulation, time policy inference, and vary the number of demonstrations to measure sample efficiency. Policies are also evaluated on a bimanual robot.
Inference Latency
ScaleVLA reads vision-language model features in one forward pass per action chunk. Inference latency is measured from sending the observation to receiving the action chunk at the client.
Sample Efficiency
The number of demonstrations per task is varied on LIBERO Goal and LIBERO Long, while training is fixed at 5,000 optimizer steps. All backbones use the same nested demonstration subsets and evaluation initial states.
ScaleVLA consistently outperforms smaller backbones in low-data regimes, with the most pronounced gains when demonstrations are scarce. With only five demonstrations per task, scaling the backbone from 2B to 552B nearly doubles the success rate on LIBERO-Goal, from 39.4% to 77.4%, while yielding substantial improvements on LIBERO-Long. Notably, ScaleVLA requires only one-fifth of the demonstrations to surpass the 2B baseline on LIBERO-Goal.
These results suggest that frontier-scale pretrained representations provide a strong foundation for action learning, enabling effective policy learning from limited robot demonstrations.

Real-Robot Experiments
To investigate whether the benefits of scaling vision-language model backbones extend beyond simulation, real-world experiments are conducted on a bimanual robot. The Plug In Charger task from RoboDojo requires sequential execution of object grasping, bimanual handover, and precise, contact-rich insertion. This task provides a challenging testbed for evaluating whether frontier-scale visual representations can benefit both multi-stage manipulation and fine-grained robotic control.
Policy robustness under spatial perturbations is also evaluated, along with the practical feasibility of deploying a frontier-scale vision-language-action policy on physical robots.

Conclusion
ScaleVLA grounds the representations of a 552B vision-language model in continuous robot actions through an action diffusion transformer trained with flow matching. Across LIBERO, LIBERO-PRO, RoboCasa, and DexJoCo, performance rises with backbone scale. On a real robot, the 552B backbone inserts a charger more reliably than the 2B backbone.
Frontier features prove most valuable where robot data is scarce, conditions shift, or contact must be precise, and they remain fast enough for subsecond control, although they require more hardware than smaller backbones. These results suggest that frontier vision-language models can serve not only as reasoning agents, but also as the representational foundation of fast, learned robot policies.
Frequently Asked Questions
What is ScaleVLA? ScaleVLA grounds vision-language model representations in continuous robot actions using an action diffusion transformer trained with flow matching.
How large is the largest backbone evaluated? The largest vision-language model backbone has 552 billion parameters.
How does scaling affect performance with few demonstrations? With five demonstrations per task, the 552B backbone achieves 77.4% success on LIBERO-Goal, compared with 39.4% for the 2B backbone.
Was ScaleVLA tested on a physical robot? Yes. It was evaluated on a bimanual robot performing the Plug In Charger task, which includes grasping, handover, and insertion.
