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

Forum topic · 小凯 · 2026-03-19

Summary

KineVLA (arXiv:2503.13845) introduces a novel kinematics-rich vision-language-action (VLA) task in which language commands densely encode kinematic attributes such as direction, trajectory, orientation, and relative displacement from initiation through completion. Unlike existing action instructions that capture kinematics only coarsely, this setting keeps task goals invariant while requiring execution trajectories to adapt to instruction-level kinematic specifications, enabling fine-grained and personalized manipulation. The authors propose KineVLA, a vision-language-action framework that explicitly decouples goal-level invariance from kinematics-level variability using a bi-level action representation and bi-level reasoning tokens, serving as explicit intermediate supervision for aligning language and actions. They also build a kinematics-aware VLA dataset spanning simulation and real robot platforms, with instruction-level kinematic variation and bi-level annotations. Experiments on LIBERO and Realman-75 robots show KineVLA consistently outperforms strong VLA baselines on kinematics-sensitive benchmarks, achieving more precise, controllable, and generalizable manipulation behavior.

Paper Overview

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

Abstract (Translation)

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) from initiation through completion, at key moments. Unlike existing action instructions that capture kinematics only coarsely or partially, this setting supports fine-grained and personalized manipulation. In this setting, task goals remain invariant while execution trajectories must adapt to instruction-level kinematic specifications.

To address this challenge, the authors propose KineVLA, a vision-language-action framework that explicitly decouples goal-level invariance from kinematics-level variability through a bi-level action representation and bi-level reasoning tokens, which serve as an explicit intermediate supervisory signal for aligning language and actions.

To support this task, the authors constructed a kinematics-aware VLA dataset covering both simulation and real robot platforms, featuring instruction-level kinematic variation and bi-level annotations.

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.

Key Contributions

A new kinematics-rich VLA task definition with dense instruction-level kinematic specifications
KineVLA framework: bi-level action representation and bi-level reasoning tokens to separate goal invariance from kinematic variability
A kinematics-aware VLA dataset spanning simulation and real robot platforms with bi-level annotations
Consistent improvements over strong VLA baselines on kinematics-sensitive benchmarks (LIBERO, Realman-75)