SIRModel: Learning Spatial Intermediate Representation to Parameter-Efficiently Fine-Tune a Vision Language Model for Manipulation
Abstract
Long-horizon robotic manipulation requires a policy to bridge task-level semantic reasoning with metric three-dimensional interaction geometry. Existing vision–language–action policies usually acquire geometry implicitly from visual tokens or introduce deterministic intermediate variables only in the image plane, which rely on expensive human annotations and training cost. This article presents a spatial Gaussian-guided hierarchical framework that uses ordered 3D Gaussian interaction regions as an explicit planning interface between vision–language reasoning and action generation. The proposed framework enables efficient adaptation of a pretrained vision–language model for robotic manipulation tasks. First, an automatic geometric enhancement pipeline converts raw robot demonstration videos into near-, mid-, and late-stage Gaussian supervision through foreground extraction, metric depth estimation, stable camera aggregation, end-effector localization, 3D lifting, and temporal grouping, without requiring manual 3D interaction annotation. The generated Gaussian representations provide structured spatial guidance, where their covariance characterizes interaction-region extent and variability rather than fully calibrated physical uncertainty. Second, a shared vision–language backbone predicts structured subtasks and Gaussian interaction regions, while a conditional diffusion executor generates future action chunks under these semantic and geometric conditions. A trajectory-to-Gaussian likelihood objective explicitly encourages consistency between generated motions and the predicted spatial interaction plan. Experiments on a mixed real-robot dataset derived from LHManip and RH20T show that our method improves trajectory tracking success from 55.7% to 70.8% over a same-backbone direct VLA baseline. Closed-loop simulation evaluation on LIBERO with 80% backbone parameter frozen achieves 85.3% average task success, demonstrating the effectiveness of explicit 3D interaction representations for spatial reasoning and long-horizon manipulation.