CoRL 2026

Equivariant Visual-Tactile Diffusion Policy for Contact-Rich Manipulation

WONG Lik Hang Kenny¹ Yiyao Ma¹ Xiu-Shen Wei²Zelong Tan¹Zhuheng Song¹Dongsheng Xie¹ Kai Chen¹ Qi Dou¹

¹ The Chinese University of Hong Kong

² Southeast University

Paper

Main Code Robot Code Dataset

Three-minute presentation03:00

01 · Overview

Learning contact-rich skills from fewer demonstrations.

VISTA combines vision, touch, and robot state in a geometry-aware policy that transfers learned behavior across spatial configurations. To our knowledge, VISTA is among the first to study workspace-level equivariance in an end-to-end visuotactile diffusion policy for contact-rich manipulation.

Abstract

Imitation learning for contact-rich manipulation requires high-quality expert data that is expensive to obtain. This makes learning a spatially robust policy under limited demonstrations a key issue. To address this, we propose VISTA, an equivariant visuotactile diffusion policy for data-efficient contact-rich imitation learning.

VISTA projects visual and tactile observations into spherical tokens, injects tactile contact cues into visual spherical directions through permutation-equivariant spherical fusion, and rotates the fused harmonic representation using the end-effector orientation. The resulting representation conditions an equivariant diffusion policy to predict spatially consistent actions. Extensive simulation and real-world experiments show substantial gains in data efficiency and spatial robustness over strong visuotactile imitation-learning baselines.

Visuo-tactile manipulationEquivariant learningDiffusion policy

80.9%Mean simulation success

+15.9 ppOver the strongest baseline

73.3%Average real-robot success

50Demonstrations per task

02 · Method

Geometry is built into the policy.

VISTA preserves how observations and actions should transform under workspace rotations, helping the policy generalize beyond the poses present in demonstrations.

Why this sequence?

Where the policy property comes from. The encoders create sensor-local spherical tokens, and fusion combines visual and tactile cues while preserving the sphere's indexing. Our experiments show that this fusion design uses both modalities effectively and improves task success. Workspace-level equivariance comes from the equivariance correction and the equivariant diffusion decoder. Following the influential ISP approach, the correction uses the end-effector rotation to align the local representation with the workspace.

Lift to the sphere

Egocentric RGB and tactile images become steerable spherical tokens with a shared geometric structure.

Fuse contact and scene cues

Permutation-equivariant fusion conditions each visual direction on tactile evidence without changing the spherical token indexing.

Predict consistent actions

The end-effector orientation aligns the fused representation to the workspace; an equivariant decoder then predicts spatially consistent actions.

03 · Experiments

Spatial robustness in simulation and on hardware.

We evaluate VISTA on eight simulated tasks and three contact-rich real-robot tasks under limited demonstrations.

Simulation

Eight geometry-sensitive tasks

50 demonstrations per task

Policy rollouts

Watch VISTA solve all eight tasks

Agent, wrist, and tactile views

Now playingGrasp Classify

Main results

VISTA leads across tasks

Data efficiency

Strong results with half the data

Ablations

All-task fusion ablation.

Simulation success rates (%) with 50 demonstrations per task and 40 evaluation rollouts.

All simulation tasks · success rate (%)
Method	Gra.	HDMI	Hole	Tube	Bot.	Can	Key	Sh.	Avg.
ISP	87.5	32.5	55.0	92.5	15.0	92.5	92.5	77.5	68.1
VISTA-early	100.0	42.5	37.5	90.0	75.0	77.5	82.5	77.5	72.8
VISTA-global-ℓ₀	100.0	17.5	52.5	90.0	77.5	77.5	57.5	82.5	69.4
ISP+Concat-Tac	100.0	37.5	67.5	85.0	67.5	92.5	77.5	65.0	74.1
VISTA (ours)	100.0	40.0	55.0	95.0	80.0	92.5	97.5	87.5	80.9

04 · Real robot

Contact-rich manipulation, up close.

Choose a task to inspect successful rollouts. Paired videos keep the external and wrist-camera views synchronized.

Surface contact

VISTA maintains stable contact while following the target erasing trajectory.

Trial 01External view

Trial 02External view

05 · Citation

Build on VISTA.

The preprint is available at arXiv:2610.03333.

BibTeX

@inproceedings{wong2026vista,
  title     = {Equivariant Visual-Tactile Diffusion Policy for
               Contact-Rich Manipulation},
  author    = {Wong, Lik Hang Kenny and Ma, Yiyao and Wei, Xiu-Shen
               and Tan, Zelong and Song, Zhuheng and Xie, Dongsheng
               and Chen, Kai and Dou, Qi},
  booktitle = {Conference on Robot Learning},
  year      = {2026},
  url       = {https://arxiv.org/abs/2610.03333}
}