[Paper Review] Imitation Learning for Object Manipulation Based on Position/Force Information Using Bilateral Control

Tsuyoshi Adachi, Kazuki Fujimoto+2 more|arXiv (Cornell University)|Nov 9, 2018

Robot Manipulation and Learning14 references4 citations

TL;DR

This paper proposes an imitation learning framework for robotic object manipulation using bilateral control to separate and utilize both position and force information from human demonstrations. By training two recurrent neural network (RNN) models—one learning position/force commands and another learning torque references—on data collected via 4-channel bilateral control, the system achieves stable, adaptive manipulation of drawing lines and curves on rulers and protractors without relying on visual sensors, demonstrating that force feedback significantly improves success rates and robustness in untrained conditions.

ABSTRACT

This study proposes an imitation learning method based on force and position information. Force information is required for precise object manipulation but is difficult to obtain because the acting and reaction forces cannnot be separated. To separate the forces, we proposed to introduce bilateral control, in which the acting and reaction forces are divided using two robots. In the proposed method, two models of neural networks learn a task; to draw a line along a ruler. We verify the possibility that force information is essential to imitate the human skill of object manipulation.

Motivation & Objective

• To address the challenge of robot adaptability in object manipulation under variable environmental conditions.
• To enable robots to learn human-like manipulation skills using both position and force feedback, which are often overlooked in conventional imitation learning.
• To overcome the limitation of force signal separation during human demonstration by employing bilateral control with master-slave robots.
• To evaluate whether force information enhances learning performance in robotic manipulation tasks beyond position and image data.
• To develop a stable and generalizable imitation learning framework that leverages force control for improved task robustness.

Proposed method

• Bilateral control is used to separate acting and reaction forces by employing a master robot (for human input) and a slave robot (for environmental interaction), enabling independent recording of force and position data.
• Two RNN models are trained: one learns position and force commands, and another learns torque reference values from the recorded bilateral control data.
• The system uses sensor data from position and six-axis force/torque sensors on both master and slave robots to capture human manipulation dynamics.
• Normalization techniques are applied to input data to improve training stability and convergence of the RNNs.
• The trained RNNs generate control commands for the slave robot to replicate the demonstrated manipulation tasks.
• Experiments are conducted on tasks such as drawing straight lines along a ruler and curves along a protractor, with no prior knowledge of the tools.

Experimental results

Research questions

• RQ1 Can force information significantly improve the success rate of robotic imitation learning in object manipulation tasks?
• RQ2 How does the integration of force and position data via bilateral control enhance the generalization of learned manipulation skills?
• RQ3 What is the comparative performance of RNN models trained on force/position commands versus torque reference values in terms of stability and success rate?
• RQ4 To what extent can a robot generalize to untrained inclinations or tools (e.g., protractors) when learning from force and position data?
• RQ5 Can stable and accurate manipulation be achieved without high-precision visual sensors by relying solely on force feedback and RNN-based control?

Key findings

• At a 15-degree inclination, the RNN model learning position and force commands (Model 1) achieved a 90% success rate, outperforming the torque-reference model (Model 2) at 75%.
• At a 45-degree inclination, both models showed comparable success rates (65% for Model 1, 70% for Model 2), indicating improved robustness at steeper angles.
• Model 2 demonstrated greater stability during failures, as it maintained control even when the task failed, unlike Model 1, which suffered from instability and divergence in torque references.
• Model 1 failed primarily due to excessive force application or torque divergence, while Model 2 failed due to loss of contact with the surface, indicating better control robustness.
• The robot successfully drew a curve using a protractor without prior training data, confirming the generalization capability of the force-based imitation learning approach.
• The trajectory from sensor data did not perfectly match the actual drawn arc, but the physical result was accurate, proving that force control alone can enable precise manipulation without high-accuracy position sensors.