OmniTacTune: Policy-Agnostic Real-World RL for Tactile Residual Adaptation of Visual Policies

Visual Policies Fail at Touch—Here's How to Fix It

Robot policies trained from human videos and demonstrations work well for basic tasks, but they struggle with contact-rich manipulation—picking fragile objects, inserting pegs, pouring liquids, anything where the robot must feel and react to forces and surfaces. The core problem: visual data alone can't capture what the robot needs to feel. Tactile sensing (pressure, vibration, contact geometry) solves this, but collecting tactile data is expensive, and tactile sensors vary wildly across hardware. OmniTacTune sidesteps this by treating tactile feedback as a correction layer bolted onto existing visual policies rather than retraining from scratch.

How It Works

The pipeline has two stages. First, the system collects rollouts—trial runs—using the pretrained visual policy on the real robot without any tactile tuning. This autonomously generates a small dataset of successes and failures. The key insight: even imperfect rollouts contain signal about where and why contact matters. The system mines these trajectories to bootstrap a tactile-aware learning phase.

Second, the robot learns a lightweight "residual policy"—a small neural network that takes tactile readings and outputs small corrections (adjustments to force, motion, or timing) to the base visual policy's actions. This residual policy is trained through online interaction: the robot attempts tasks, collects real tactile data during execution, and refines the correction layer over 40–80 minutes of real-world trial-and-error.

The design is policy-agnostic: it doesn't assume anything about the visual policy's architecture or training procedure, so it works with policies from videos, teleoperation, or robot demonstrations. It's also sensor-agnostic—the tactile representation (raw pressure arrays, learned embeddings, etc.) is abstracted away, so the same pipeline adapts to different tactile hardware.

Why It Matters

Visual policies are scalable; tactile data collection is not. By using tactile as a residual adaptation layer—a lightweight patch that learns what the vision-based controller misses—the authors avoid the cost of collecting large tactile datasets while gaining the benefits of tactile feedback. The results are striking: on four real-world contact-rich tasks (insertion, pouring, grasping, pushing), visual-only policies succeeded 5–40% of the time. After OmniTacTune's two-stage adaptation, success rates jumped to 85–100%.

The practical payoff is speed. Forty to eighty minutes of online learning on the actual robot is fast by real-world RL standards, and the method generalizes across tasks, base policies, and tactile sensors—a rare property in tactile manipulation research, where generalization typically breaks when you switch hardware or task. This makes the approach viable for practitioners who want to upgrade existing visual policies without rebuilding from tactile up.