Sensorized gripper for human demonstrations

TL;DR

This work tackles the challenge of easy robot programming in unstructured environments by recording human demonstrations with a sensorized gripper and replaying them on a 7-DoF robot using impedance control. It combines a Cartesian motion-generation framework that jointly computes the joint-space trajectory and optimizes the base pose to maximize manipulability, with Gaussian Mixture Regression extracting a representative demonstration from brief trials to drive replay. The key contributions are the sensorized gripper design, base-location optimization, and impedance-based replay, along with a quantified assessment of haptic mismatch between human demonstrations and robot execution. The approach demonstrates that off-the-shelf hardware can enable rapid, safe, and repeatable programming for manipulation tasks like box-in-box assembly, using around 100 seconds of demonstration data.

Abstract

Ease of programming is a key factor in making robots ubiquitous in unstructured environments. In this work, we present a sensorized gripper built with off-the-shelf parts, used to record human demonstrations of a box in box assembly task. With very few trials of short interval timings each, we show that a robot can repeat the task successfully. We adopt a Cartesian approach to robot motion generation by computing the joint space solution while concurrently solving for the optimal robot position, to maximise manipulability. The statistics of the human demonstration are extracted using Gaussian Mixture Models (GMM) and the robot is commanded using impedance control.

Figures (5)

• Figure 1: a) The tool is the only common aspect in the task. b) We choose off-the-shelf gripper and bicycle brake handle for the actuator lever to assemble the tool.
• Figure 2: a) For robot base frame scenario $\{s\}$ varying in a plane, robot manipulability is evaluated. b) Robot joint state is computed from gripper marker data using first order dynamics. c) The optimal robot base frame is determined after optimising the integral of manipulability across all scenarios.
• Figure 3: Statistics of human demonstration using sensorized gripper.
• Figure 4: a) Optimal robot base location computed over a large region surrounding the demonstration. b) Impedance control.
• Figure 5: (a) Task haptics were sensed at the receptacle frame $\{R\}$, (b) In this work, we deliberately vary the initial location of the box to record the resulting haptic mismatch in case of in-hand pose uncertainty, (c) Haptic mismatch, i.e., the difference in demonstration forces and robot forces as measured in the receptacle frame $\{R\}$.