A two-stage training framework introduces a dedicated action prior learning stage before vision-language-action (VLA) training. By training a flow-matching-based encoder-decoder action module solely on unconditioned action trajectories, we equip the action head with structured knowledge of the action distribution before it encounters visual-language inputs. An early-stage latent alignment distillation strategy further accelerates cross-modal convergence by anchoring the VLM to the learned action embedding space. Experiments across 13 cross-embodiment tasks validate that the proposed approach yields faster convergence, higher success rates, and effective history compression at negligible cost. We further demonstrate that enriching the Stage 1 data produces a stronger action prior that directly transfers to improved downstream VLA performance, confirming the scalability of the proposed framework.
Action Prior Learning
Overall, these approaches show that latent action or dynamics representations can improve robot policy learning. Most of them, however, start from visual data, future-state prediction, or vision-action policy learning. Our setting instead uses robot action trajectories directly, without visual observations or language instructions. This isolates the intrinsic statistics of physical motion and keeps the learning stage lightweight.
Analysis of Training Dynamics and Optimization Stability
Beyond final success rates and rollout behavior, we further examine how action prior learning changes the early optimization process of VLA training. The key question is whether Stage 1 action prior learning can reduce the burden of Stage 2 VLA training, where the model must align visual-language features with continuous action generation. Experiments compare the first 3,000 Stage 2 steps of three variants under the same data, hyperparameters, and VLM backbone, isolating the effect of action prior initialization and history conditioning.
Stable initialization. The no-prior model starts from a poorly calibrated action module and suffers a much larger initial prediction loss. In contrast, prior-based variants begin from a decoder that already models feasible motion, reducing the initial loss by nearly a significant margin and reaching low loss much earlier. This suggests that Stage 1 handles the low-level action calibration before cross-modal alignment begins. Stage 2 can therefore focus less on discovering valid motion patterns from scratch and more on adapting these patterns to visual-language conditions.
Smoother gradients. Random action initialization also produces large and recurrent gradient spikes, while the learned decoder keeps gradients small and stable from the beginning. This matters because stable action gradients reduce interference with the already useful representations in the VLM backbone.
Faster validation convergence. The MAE curve confirms that the gain is not limited to training loss. Prior-based variants reach the same validation error much earlier, indicating that the learned action prior also serves as an optimization prior for cross-modal policy learning. The history variant further reaches the lowest MAE after the initial transient, consistent with its stronger downstream performance.
Conclusion
We presented a two-stage training framework that introduces a dedicated action prior learning stage before VLA training. By training a flow-matching-based encoder-decoder action module solely on unconditioned action trajectories, we equip the action head with structured knowledge of the action distribution before it encounters visual-language inputs. An early-stage latent alignment distillation strategy further accelerates cross-modal convergence by anchoring the VLM to the learned action embedding space. Experiments across 13 cross-embodiment tasks validate that the proposed approach yields faster convergence, higher success rates, and effective history compression at negligible cost. We further demonstrate that enriching the Stage 1 data produces a stronger action prior that directly transfers to improved downstream VLA performance, confirming the scalability of the proposed framework.
Frequently Asked Questions
What is action prior learning and how does it differ from other approaches? Action prior learning trains a module on robot action trajectories alone, without using visual data or language instructions. This isolates the statistics of physical motion, keeping the learning lightweight compared to methods that start from visual observations.
How does the two-stage training framework improve VLA policy learning? Stage 1 calibrates the action decoder on motion patterns before Stage 2 cross-modal training begins. This provides stable initialization, smoother gradients, and faster validation convergence, reducing the burden of learning action generation from scratch.
Does enriching Stage 1 data improve downstream performance? Yes, experiments confirm that adding more action data in the action prior learning stage produces a stronger prior, which directly transfers to higher success rates in the full VLA model.
What is the role of history conditioning in the framework? The history variant of the action prior module reaches the lowest validation error after initial training, and demonstrates stronger downstream performance, showing that modeling temporal context in action sequences provides additional benefits.