Predicting object properties based on movement kinematics.

Area of Science:

• Robotics
• Human-Computer Interaction
• Biomechanics

Background:

• Grasping and transporting objects requires scaling grip and load forces based on object properties like weight.
• Current methods for weight estimation, such as image recognition in robots, often depend heavily on prior learning.
• A less experience-dependent approach is needed for robots to handle a wider variety of objects.

Purpose of the Study:

• To evaluate the feasibility of predicting an object's weight class using upper body kinematics or object velocity profiles.
• To investigate the impact of time series length and cross-validation procedures on prediction accuracy.
• To develop a robot weight estimation method less reliant on prior learning.

Main Methods:

• Recorded movement kinematics of 12 participants performing a replacement task with objects of varying, unknown weights.
• Utilized an optical motion tracking system to capture upper body angles.
• Applied discrete cosine transform for time series smoothing/compression and support vector machine for supervised weight class prediction.

Main Results:

• Achieved good prediction accuracy for object weight class, influenced by cross-validation procedures and time series length.
• Reliable weight prediction was possible early in the movement (within 300 ms).
• The proposed method demonstrated effectiveness in estimating object weight without extensive prior learning.

Conclusions:

• Predicting object weight class from upper body kinematics during a replacement task is feasible.
• The approach offers a promising alternative to traditional, experience-dependent weight estimation methods for robots.
• This technique enhances robotic capabilities for grasping and manipulating diverse objects.