Where Should Action Generation Begin? A Learnable Source Prior for Generative Robot Policies

A research team has introduced LeaP, a learnable source prior that replaces the standard Gaussian distribution used to initialize action generation in robot policies, reporting an average success rate of 81.6% across 15 RoboTwin manipulation tasks [1]. Generative robot policies conventionally draw initial action samples from an observation-independent standard Gaussian distribution, a design choice that the authors argue has been largely unexamined [1]. The LeaP framework substitutes that fixed noise source with a proprioception-conditioned diagonal Gaussian parameterized by a lightweight multi-layer perceptron. The model jointly predicts the mean and state-adaptive variance of the source distribution while leaving the downstream generator architecture and inference solver unchanged [1]. This observation-informed but stochastic initialization allows the generator to concentrate on precise action refinement rather than transporting samples from an uninformed starting point [1]. On 15 RoboTwin manipulation tasks, LeaP outperformed four representative baselines—including deterministic-source methods, a no-prior counterpart, and a diffusion-bridge policy—by margins ranging from 6.5 to 25.5 percentage points [1]. The same prior consistently improved both flow-matching and diffusion-bridge generators while using fewer parameters and converging faster [1]. The performance advantage carried over to real-world deployment, where LeaP attained the best results among tested methods [1]. The findings suggest that the source distribution constitutes an independent and reusable design axis for generative robot policies, complementary to the choice of generative dynamics [1]. The work arrives amid a broader surge in generative AI research that has accelerated since the introduction of the transformer architecture in 2017, which enabled the rapid scaling of large language models and fueled exponential investment in the field [5]. As AI systems increasingly influence or automate human decision-making, questions of algorithmic bias, transparency, and accountability have gained prominence in parallel with technical advances [4]. The paper was submitted to arXiv on June 16, 2026, and is available through the arXivLabs framework, which allows community collaborators to develop and share new features on the platform [1].