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

2026-06-15

AI Summary

This work addresses the limitations of conventional generative robot policies that employ a standard, observation-agnostic Gaussian distribution as the action prior, which constrains performance. The authors propose LeaP—a learnable, proprioception-conditioned diagonal Gaussian prior—that leverages a lightweight MLP to jointly predict state-adaptive means and variances, thereby providing a superior initialization for action generation. Notably, this approach treats the source distribution as an independent and reusable design dimension for the first time, significantly enhancing performance without altering the downstream generator architecture. Evaluated across 15 RoboTwin tasks, LeaP achieves an average success rate of 81.6%, outperforming four baseline methods by 6.5–25.5 percentage points while using fewer parameters and converging faster. Its efficacy is further validated on a physical robot platform.

Abstract

Generative robot policies typically begin action generation from an observation-independent standard Gaussian distribution, leaving the choice of source distribution underexplored. This work asks a simple question: where should action generation begin? We propose LeaP, a Learnable source Prior that replaces the standard Gaussian with a proprioception-conditioned diagonal Gaussian over action chunks. Parameterized by a lightweight MLP, LeaP jointly predicts the mean and state-adaptive variance of the source distribution, while keeping the downstream generator architecture and inference solver unchanged. This design provides an observation-informed yet stochastic initialization, allowing the generator to focus on precise action refinement rather than transporting samples from an uninformed noise source. On 15 RoboTwin manipulation tasks, LeaP achieves an average success rate of 81.6%, outperforming four representative baselines -- including deterministic-source methods, a no-prior counterpart, and a diffusion-bridge policy -- by 6.5 to 25.5 percentage points. The same prior consistently improves both flow-matching and diffusion-bridge generators, while using fewer parameters and converging faster. The advantage carries over to real-world deployment, where LeaP attains the best performance. These results suggest that the source distribution is an independent and reusable design axis for generative robot policies, complementary to the choice of generative dynamics.

Problem

• generative robot policies
• source distribution
• action generation
• learnable prior
• proprioception-conditioned

Innovation

• Learnable Prior
• Generative Robot Policies
• Source Distribution
• Proprioception-Conditioned
• Action Generation