Publikation
CVLAMP: Compliant Vision Language Action Movement Primitives
Vamsi Krishna Origanti; Alexander Fabisch; Frank Kirchner
In: Late-Breaking Result, 2026 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS-2026), September 27 - October 1, Pittsburgh, PA, USA, IEEE/RSJ International Conference on Intelligent Robots and Systems, 9/2026.
Zusammenfassung
We present Compliant Vision-Language-Action Movement Primitives (CVLAMP), a closed-loop framework for contact-rich dual-arm manipulation. CVLAMP conditions a VLA backbone on images, language, Cartesian state, and force history, decoding a compact Cartesian motion primitive in place of a dense action chunk. Each primitive couples a Dynamic Movement Primitive (DMP) pose rollout with shared-phase wrench and gripper profiles, executed by a phase-governed controller that can run, slow, hold, or abort from real-time feedback. The architecture extends a pretrained π0.5 flow-matching VLA with cross-attention-fused state and force encoders, trained with a composite regression, rollout-consistency, and mode-classification objective via staged, single-GPU LoRA adaptation of the released checkpoint. In simulation, the model localizes the goal object from vision alone with a mean 2D position error of approximately 5.5 cm on held-out, previously unseen object positions, and the closed-loop system reliably approaches and grasps the target object; full task completion and real-robot deployment are ongoing work.