KineVLA: Kinematics-Aware Vision-Language-Action Models for Fine-Grained Robot Manipulation

Paper Overview

• Field: Computer Vision (CV)
• Authors: Gaoge Han, Zhengqing Gao, Ziwen Li
• Published: 2025-03-18
• arXiv: 2503.13845

Summary

This paper introduces a novel kinematics-rich vision-language-action (VLA) task, in which language commands densely encode diverse kinematic attributes—such as direction, trajectory, orientation, and relative displacement—at key moments from initiation through completion. Unlike existing action instructions that capture kinematics only coarsely or partially, this formulation supports fine-grained and personalized manipulation.

In this setting, task goals remain invariant while execution trajectories must adapt to instruction-level kinematic specifications.

Proposed Approach

To address this challenge, the authors propose KineVLA, a vision-language-action framework that:

• Explicitly decouples goal-level invariance from kinematics-level variability
• Uses a bi-level action representation and bi-level reasoning tokens as explicit intermediate supervision to align language and action

Dataset

The authors construct kinematics-aware VLA datasets covering both simulation and real robot platforms, featuring:

• Instruction-level kinematic variations
• Bi-level annotations

Results

Extensive experiments on LIBERO and Realman-75 robots demonstrate that KineVLA consistently outperforms strong VLA baselines on kinematics-sensitive benchmarks, achieving more precise, controllable, and generalizable manipulation behavior.

---

*Auto-collected on 2026-03-19.*