Abstract
Teaching robots complex manipulation skills through Learning from Demonstration (LfD) often requires expensive, proprietary hardware. This paper introduces a low-cost (~250 EUR), open-source sensor glove equipped with force feedback for teleoperation. By mapping human finger movements to a humanoid robot and providing tactile feedback via vibration motors, the system facilitates the collection of high-quality demonstration data. To handle the inherent noise in human demonstrations, we employ a probabilistic movement representation using Bayesian linear regression. This allows the robot to learn a generalized control strategy from as few as two demonstrations. We demonstrate the system's efficacy by teaching an iCub humanoid robot a cup-stacking task.
Introduction
Learning from Demonstration is a powerful paradigm for transferring human skills to robots. However, two significant challenges persist: the high cost of haptic teleoperation hardware and the difficulty of extracting robust, executable policies from noisy human data. Commercial sensor gloves with force feedback are often prohibitively expensive, limiting accessibility for researchers. Furthermore, direct playback of teleoperated trajectories often fails due to inconsistencies in human movement. This work addresses these issues by providing an affordable hardware solution and a probabilistic framework for trajectory learning.
System Architecture
The hardware core is a custom-built sensor glove designed for approximately 250 EUR. It utilizes an Arduino Mega 2560 to interface with 4.5-inch flex sensors that measure finger flexion. To close the loop with the operator, VPM2 coin vibration motors are integrated into the fingertips. These motors provide proportional tactile feedback based on the robot's own tactile sensor readings, allowing the operator to "feel" contact forces and avoid damaging fragile objects.
The glove streams data at 350Hz to a host computer. For spatial tracking, an external motion capture system currently monitors the operator's wrist position, which is then translated into robot arm movements via inverse kinematics.
Probabilistic Trajectory Learning
To transform raw teleoperation data into a robust robotic policy, we model demonstrations using probabilistic trajectory representations. Rather than simple averaging, we apply Bayesian linear regression with Gaussian basis functions to the state vectors. This approach learns a distribution over the demonstrated trajectories, capturing both the mean path and the associated variance.
During autonomous reproduction, the robot follows the mean of this learned distribution using a low-level impedance controller. This probabilistic modeling naturally filters out human-induced noise and provides a framework for future adaptive control based on learned variance.
Experimental Evaluation
The system was validated using the iCub humanoid robot in a cup-stacking task. The operator used the glove to teleoperate the robot, stacking one plastic cup onto another. The tactile feedback was found to be critical for maintaining a secure grip without crushing the lightweight cups.
Remarkably, the probabilistic model was able to learn the 13-joint manipulation task effectively using only two human demonstrations. The resulting autonomous execution was smooth and reliable, demonstrating the system's data efficiency and robustness.
Conclusion and Future Work
We have presented an affordable, open-source platform for haptic teleoperation and imitation learning. Our results show that high-fidelity teleoperation does not require expensive hardware and that probabilistic modeling can effectively generalize from sparse, noisy demonstrations.
Future work will focus on increasing the glove's portability by replacing external motion capture with onboard IMUs. Additionally, we aim to implement adaptive control strategies that utilize the learned variance—allowing the robot to be stiff where precision is required and compliant where demonstrations show high variability.
Summary updated