Foundation model manipulation · tactile adaptation

ViTaR: Visuo-Tactile Residual Adaptation for Foundation VLA Manipulation

Touch adapts execution without rewriting a frozen VLA policy.

PaperPDF CodeComing soon WebsiteProject page VideoProject video

01 / Overview

Touch as a bounded execution modulator.

ViTaR preserves a frozen VLA’s semantic action, then uses local tactile evidence for conservative execution calibration.

ViTaR project video preview showing three real-world robot manipulation tasks

ViTaR teaser figure showing tactile residual adaptation over a frozen VLA policy.

ViTaR preserves the semantic action proposed by a frozen VLA, and only injects a tactile-conditioned residual when local contact evidence calls for calibration.

How it works

A two-stage design determines whether a correction is locally justified, then selects and scales a structured residual before execution.

ViTaR framework overview: a frozen VLA is augmented with Effect-Guided Modeling and Residual Action Modulation.

Frozen foundation VLA

Language, RGB, and proprioception produce a reference action while the pretrained policy remains intact.

Effect-Guided Modeling

Outcome-grounded preference evidence identifies whether and which correction is locally useful.

Residual Action Modulation

Visuotactile observations select and continuously scale a bounded residual for execution.

02 / Results

Improved success in simulation and on the physical robot.

Across contact-rich tasks, ViTaR lifts a frozen base policy through conservative, tactile-conditioned corrections.

UniVTAC average success

61.3%+30.6 percentage points over frozen OpenVLA-OFT

Physical-robot average success

48.3%+30.0 percentage points over frozen OpenVLA-OFT

Design principle

Retain pretrained semantic intent; let touch make only the local correction.

Physical-robot tasks

Better contact calibration carries into the real world.

ViTaR is evaluated on precision insertion, stable grasping, and sustained sliding contact with a RealMan RM65-B robot and a tactile parallel gripper.

Open chart PDF ↗

Physical robot success rates for Insert Tube, Lift Bottle, Wiping the Board, and average across four methods.

03 / Real-world experiments

ViTaR in physical contact-rich manipulation.

Three representative real-robot tasks demonstrate precision alignment, secure grasping, and sustained contact.

Precision alignment

Insert Hole

Fine contact adjustment for insertion.

Stable grasping

Lift Bottle

Contact-aware grasp stabilization.

Sustained sliding contact

Wiping the Board

Adaptive force-sensitive motion during contact.