Non-prehensile manipulation offers a robust alternative to traditional pick-and-place methods for object repositioning. However, learning such skills with dexterous, multi-fingered hands remains largely unexplored, leaving their potential for stable and efficient manipulation underutilized. Progress has been limited by the lack of large-scale, contact-aware non-prehensile datasets for dexterous hands and the absence of wrist-finger control policies.
To bridge these gaps, we present DexMove, a tactile-guided non-prehensile manipulation framework for dexterous hands. DexMove combines a scalable simulation pipeline that generates physically plausible wrist-finger trajectories with a wearable device, which captures multi-finger contact data from human demonstrations using vision-based tactile sensors. Using these data, we train a flow-based policy that enables real-time, synergistic wrist-finger control for robust non-prehensile manipulation of diverse tabletop objects.
In real-world experiments, DexMove successfully manipulated six objects of varying shapes and materials, achieving a 77.8% success rate. Our method outperforms ablated baselines by 36.6% and improves efficiency by nearly 300%. Furthermore, the learned policy generalizes to language-conditioned, long-horizon tasks such as object sorting and desktop tidying.
Figure 1: Overview of DexMove. The framework integrates synthetic non-prehensile manipulation trajectories and human-demonstrated tactile data to train a flow-matching policy for dexterous hands. The learned policy generalizes across diverse objects, surface frictions, and various language-conditioned tasks such as tidying.
Figure 2: Trajectory data verification and force augmentation. (a,b) After generating initial grasp poses, the set of target directions along which the object can be manipulated is pruned using physics simulation, and the fingertips' trajectories are computed by sampling target object poses. (c) The contact force is synthesized as a surrogate for the finger’s penetration depth into the object, and we augment the trajectories with diverse forces.
Figure 3: A wearable device for collecting real tactile-force data. The collected force during manipulation is characterized by the displacement of markers within the vision-based tactile sensor.
The tactile sensor is based on the PP-Tac design, which has been open sourced. We add some markers on the gel surface to track the contact deformation.
@inproceedings{DexMove_ICLR2026,
title={DexMove: Learning Tactile-Guided Non-Prehensile Manipulation with Dexterous Hands},
author={Pei, Lin and Yuzhe, Huang and Wanlin, Li and Chenxi, Xiao and Ziyuan, Jiao},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=dT3ZciXvNX}
}