💻 computer science

HapCompass: A Rotational Haptic Device for Contact-Rich Robotic Teleoperation

The paper introduces HapCompass, a low-cost wearable haptic device that mechanically rotates a single actuator to provide intuitive 2D directional cues, significantly improving teleoperated manipulation performance and imitation learning outcomes compared to vision-only and non-directional feedback baselines.

Original authors: Xiangshan Tan, Jingtian Ji, Tianchong Jiang, Pedro Lopes, Matthew R. Walter

Published 2026-04-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Xiangshan Tan, Jingtian Ji, Tianchong Jiang, Pedro Lopes, Matthew R. Walter

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to thread a needle while wearing thick, clumsy winter gloves, and you can only see the needle through a tiny, foggy keyhole. That is roughly what it feels like to control a robot arm to do delicate tasks, like plugging in a USB drive or inserting a key into a lock, using only your eyes. You can see where the robot is, but you can't feel when it bumps into something or if it's pushing too hard.

This is the problem HAPCOMPASS solves.

The Problem: The "Silent" Robot

In the world of robotics, there is a big gap between the robot and the human. The robot has super-sensitive "fingers" that can feel exactly how hard it's pushing and in which direction. But the human operator? They are often flying blind, relying only on a video feed.

Existing solutions try to fix this by making the controller vibrate (like a game controller buzzing when you hit a wall). But that's like a car horn that just says "BEEP!" It tells you something happened, but it doesn't tell you which way to turn the steering wheel to avoid the crash. Other high-tech gloves are too expensive, bulky, or confusing because they vibrate in too many places at once, creating a "static noise" in your brain.

The Solution: The Haptic Compass

The researchers built HAPCOMPASS, a wearable device that fits on your finger like a ring. Think of it as a mechanical compass for your sense of touch.

Here is the clever trick:

One Vibrator, Many Directions: Instead of having a dozen tiny motors, HAPCOMPASS uses just one strong vibration motor (an LRA).
The Spinning Trick: This motor is mounted on a tiny, spinning platform. If the robot hits a wall on the left, the device spins the motor to the left and vibrates. Your brain feels a "pull" to the left. If the robot is pushing up, the motor spins up and vibrates.
The Result: It's like having a tiny, invisible hand on your finger gently tugging you in the exact direction the robot needs to move to avoid a crash. It turns a vague "buzz" into a clear, directional arrow.

How It Works in Real Life

The team tested this on three tricky tasks:

The Key: Putting a fragile wooden key into a tight lock.
The USB: Plugging a USB drive into a computer port without bending it.
The Spaghetti: Trying to push a raw spaghetti noodle into a jar full of rice and blocks without snapping it.

The Results:

Without HAPCOMPASS: People struggled. They broke keys, bent USB drives, and snapped spaghetti because they couldn't feel the resistance.
With HAPCOMPASS: It was like suddenly gaining a sixth sense. Operators finished tasks faster, broke fewer things, and applied much gentler force. They could "feel" the friction and adjust their grip instantly, just like a human would if they were holding the object directly.

The Bonus: Teaching Robots to Learn

The researchers also tested if this device could help robots learn on their own. They recorded humans doing the tasks with and without the device, then taught a robot AI to copy them.

The AI trained on "blind" humans (no haptics) was clumsy and often broke the objects.
The AI trained on "HAPCOMPASS" humans learned much faster and became much more careful. It seems that when humans can feel the task, they move more smoothly, creating better "textbooks" for the robot to study.

Why This Matters

HAPCOMPASS is a low-cost, simple, and brilliant way to bridge the gap between human intuition and robotic precision. It turns a frustrating, visual-only game of "guess and check" into a smooth, intuitive dance where the operator can actually feel what the robot is doing.

In short: It gives robots a voice, and gives humans a way to hear it through their fingertips.

1. Problem Statement

Robotic teleoperation for contact-rich manipulation tasks (e.g., tight-clearance insertions like USB or key insertion, or navigating granular media) is hindered by a sensory asymmetry between the human operator and the robot.

The Gap: While robots are equipped with advanced tactile and force sensors, human operators often lack corresponding channels to perceive contact interactions during teleoperation.
Limitations of Current Solutions:
- Vision-only interfaces: Force operators into low-bandwidth, vision-dominant loops, leading to clumsy movements and inefficient data for imitation learning.
- Non-directional vibrations (e.g., VR controllers): Provide intensity cues but fail to convey the direction of contact forces, which is critical for fine-grained adjustments.
- Haptic gloves: Typically render only normal (grasping) forces, lacking tangential force directionality.
- Vibrotactile arrays: Can suffer from perceptual interference when multiple actuators vibrate simultaneously.
- Bilateral teleoperation (grounded devices): High fidelity but expensive, non-portable, and restrict the operator's workspace.

2. Methodology: The HAPCOMPASS System

The authors propose HAPCOMPASS, a novel, low-cost wearable haptic device designed to render 2D directional cues mechanically.

A. Hardware Design

Core Mechanism: Instead of using multiple actuators, HAPCOMPASS uses a single Linear Resonant Actuator (LRA) mounted on a rotating platform.
Principle of Operation:
1. The LRA is driven by an asymmetric waveform, creating a perceived "pulling" sensation along its actuation axis.
2. A servo motor rotates the LRA to orient this axis in the x-y plane.
3. By changing the orientation of the LRA, the device renders directional haptic cues corresponding to the direction of contact forces.
Form Factor: A ring worn on the middle finger, allowing the index finger and thumb to rest near the actuator.
Isolation: A decoupled two-part housing with O-rings isolates the vibration from the main body to reduce noise and ensure the user feels only the intended directional pull.
Cost: Approximately $140 USD.

B. System Architecture

The teleoperation system integrates visual and haptic feedback:

Command Path: Uses Meta Quest 3 for hand tracking. Human hand poses are retargeted to robot end-effector poses.
Feedback Path:
- Visual: Video passthrough + wrist-mounted camera.
- Haptic: Data is sourced from tactile sensors on the robot's gripper fingers and a wrist-mounted Force-Torque (F/T) sensor.
Rendering Algorithm:
- Calculates the change in tactile force vector ( $\Delta F_{sensor}$ ) relative to a baseline.
- Projects the 3D force vector onto a 2D plane relevant to the task (e.g., insertion axis).
- Mapping: The vector's direction controls the servo motor's angle ( $\theta$ ), and the vector's magnitude controls the LRA's vibration amplitude ( $A$ ).

3. Key Contributions

Novel Device: HAPCOMPASS, a low-cost wearable that renders 2D directional cues by mechanically rotating a single LRA, avoiding the complexity of multi-actuator arrays.
Integrated System: A full teleoperation stack combining vision-based hand tracking with directional haptic feedback for contact-rich tasks.
Perception Validation: Quantitative evidence that humans can reliably interpret directional cues from the device.
Teleoperation Performance: Demonstration that directional feedback significantly improves task success, reduces completion time, and lowers contact forces compared to vision-only and non-directional baselines.
Imitation Learning Impact: Preliminary evidence that demonstrations collected with HAPCOMPASS yield higher-quality training data, resulting in more robust learned policies.

4. Experimental Results

The system was evaluated through three experiments involving 12 participants across three tasks: Key Insertion, USB Insertion, and Spaghetti Probing.

A. Directional Rendering (Experiment 1)

4-AFC Task (Cardinal directions): Participants achieved 84.4% accuracy (mean angular error: 24.4°).
8-AFC Task (Cardinal + Intercardinal): Participants achieved 67.2% accuracy (mean angular error: 28.4°), significantly above the 12.5% chance level.
Error Analysis: Confusion was highest between "Up" and "Down" due to mechanical damping from the drive belt, and "Down-Left" was often confused with "Left."

B. Teleoperation Performance (Experiment 2)

Comparing four conditions: Vision-Only (C1), Vision + Non-directional (C2), Vision + Controller Vibration (C3), and Vision + Directional (C4).

Success Rates:
- Key Insertion: C4 achieved 100% success, compared to 63.2% (C1) and 73.7% (C2).
- USB Insertion: C4 and C3 both achieved ~98.3%, outperforming C1 (93.3%).
- Spaghetti Probing: C4 achieved the highest success rate (60.0%).
Safety & Efficiency:
- Contact Force: C4 resulted in the lowest maximum contact forces in Key and USB insertion tasks.
- Bending Torque: C4 produced significantly lower bending torque in Key Insertion, indicating safer, more aligned insertions.
- Time: While C3 (controller-based) was fastest due to stable tracking, C4 was the most efficient among vision-based hand-tracking conditions.

C. Imitation Learning (Experiment 3)

Policies trained on demonstrations collected with HAPCOMPASS (C4) achieved a 90% success rate on the Key Insertion task.
The baseline policy (trained on C3 data) achieved only 60%.
The C4-trained policy exhibited reactive adjustments to contact forces, whereas the C3 policy often failed due to excessive force.

5. Significance and Future Work

Significance:

HAPCOMPASS bridges the gap between high-fidelity bilateral teleoperation and flexible, portable unilateral interfaces.
It proves that directional haptic feedback is superior to non-directional intensity cues for tasks requiring fine motor control and force regulation.
It enhances the quality of datasets for Imitation Learning, suggesting that human operators can demonstrate "safer" and more precise behaviors when they can "feel" the direction of contact forces.

Limitations & Future Directions:

Current Limitations: Restricted to 2D directional cues (not 3D); some end-to-end latency remains; evaluation limited to specific tasks.
Future Work: Integrating actuators for normal-force cues, conducting systematic studies on how haptic feedback affects learning datasets, and generalizing to more diverse manipulation scenarios.

Conclusion:
HAPCOMPASS offers a practical, low-cost solution to the "sensory asymmetry" problem in robotics, enabling safer, more successful teleoperation and generating higher-quality data for training autonomous robotic policies.