Contact-Grounded Policy — Dexterous Visuotactile Policy with Generative Contact Grounding

Venue: RSS 2026 (Manipulation session) · Authors: Zhengtong Xu (Purdue), Yeping Wang (UW–Madison), Ben Abbatematteo, Jom Preechayasomboon, Sonny Chan, Nicholas Colonnese, Amirhossein H. Memar — Purdue × Meta Reality Labs Research × UW–Madison · arXiv: 2603.05687 · project · also Outstanding award, CVPR 2026 Sense-of-Space workshop Category: Visuo-tactile dexterous policy Trend tag: RSS 2026 thread 4 — contact as representation

Compiled from the verified RSS 2026 abstract and the paper's Fig. 1.

Key figure

Figure 1 of the paper. Left: the two demonstration channels — virtual teleoperation with simulated visuotactile observations (dense whole-hand tactile arrays on the Tesollo DG-5F) and physical teleoperation with real observations (Allegro V5 + four Digit360 fingertip sensors); below, the CGP pipeline — the policy predicts tactile, actual robot state, and target robot state jointly, and the contact-consistency map turns them into compliance-controller targets. Right: the task suite — fragile egg grasping, in-hand box flipping, and dish wiping in sim; jar opening and real in-hand box flipping on hardware. The predicted-tactile-to-target mapping in the center strip is the paper's thesis in one image: the policy commands contacts, not just poses.

Problem

Multi-finger task success hinges on continuously evolving multi-point contacts that are hypersensitive to geometry, friction transitions, and slip. Most tactile policies treat touch as an extra observation — they model neither the contact state itself nor how policy outputs interact with the low-level controller.

Method

Two coupled components:

1. Conditional diffusion model forecasting coupled trajectories of actual robot state + tactile feedback in a compressed latent space — the policy predicts what it will feel, not just where it will move.
2. Learned contact-consistency mapping converting each predicted state–tactile pair into executable targets for a compliance controller, so the controller realizes the intended contacts rather than just the intended poses.

Results (as reported)

• Hardware: four-finger Allegro V5 + Digit360 fingertips; simulation: five-finger Tesollo DG-5F with dense whole-hand tactile arrays.
• Outperforms visuomotor and visuotactile diffusion-policy baselines across in-hand manipulation, delicate grasping, and tool use.

Significance

The most explicit "close the loop through contact" design at RSS 2026: prediction of feeling, then a learned map from predicted feeling to controller targets. Together with ViTacFormer (predictive tactile representation) and Force Policy (#128, hybrid force–position in the interaction frame), it marks the session-level shift from touch-as-input to touch-as-predicted-state — the tactile mirror of the world-model thread. See Review-Tactile-VLA.

← RSS 2026 survey · Home