Design Principles for Multi-Axis, Large Force Detecting, and Slip-Predicting Sensor Arrays for Use in Robotic Applications

ResearchPoster-Chuah-v2.pdf

Having a complete picture of the ground interactions of a legged robot is a necessity in enabling high-speed, dynamic, ground locomotion. It is especially important for high speed ground locomotion as well as in aggressive and dynamic movements that we maximize traction (i.e. go right to the edge of the friction cone) and risk slipping to turn around, or accelerate and run at top speed. Hence, we need to know how much normal and shear force we have applied, and how far we can go before our foot loses traction. In contrast with dexterous manipulation, we only want to keep the object in our grasp from slipping, so we can be more conservative and apply larger normal forces. Current force sensing methods such as using load cells are inadequate to address the unique demands of robotic legged locomotion, and are vulnerable to inertial noise upon high acceleration. 

Using new design principles and methodologies, we have developed a multi-axis, large force detecting foot sensor for legged robots. It is a monolithic composite structure that is composed of a piezoresistive sensor array PCB completely embedded in a protective polyurethane rubber layer. The composite architecture allows for compliance and traction during ground contact, while deformation alters the measured stress distribution. Using machine learning, we are able to map the local sampling of stress inside the polymeric footpad to forces in three axes with high accuracy. This footpad sensor is intended for use on the MIT Cheetah to provide a complete picture of the ground interaction forces that is a necessity in enabling high-speed and dynamic ground locomotion.

Motivation

Human beings and animals are capable of great feats of dynamic locomotion, even when the ground interactions are uncertain. A great example of this is of tennis players intentionally sliding on clay courts to reach the ball. These tennis players are able to predict the amount that they will slide by and that they would stop in the perfect position to intercept the ball with their tennis racket.

So why are humanoid robots unable to carry out such actions? It is not a limitation on the actuation side. It has long been acknowledged that motors and hydraulic actuators are capable of greater force/torque output than Olympic weightlifters. Rather, there is a need for a sensor that is able to capture the large normal and shear forces, that is robust to impacts, and that is able to relay the data in real-time in order to allow a robot to predict its motions with certainty.

Human beings have such sensors in abundance in our skin, in the form of mechanoreceptors that allow us to sense pressure, vibrations and forces. This enables us to get a complete picture of the ground interactions at each footfall, and predict our next steps with certainty in spite of changing terrain conditions. To empower the next generation of legged robots, we would need to develop similar multi-axis force sensors. With the capability to measure normal and shear forces in real-time, robots would then be able to predict if they are about to slip and fall, and take preventive measures to safeguard themselves.

There is a need to develop a novel multi-axis force sensor as current force sensing methods are inadequate to address the unique demands of legged locomotion. Typically, most robots use load cells or force/torque sensors in the feet to sense the ground contact interactions. These strain gauge based load cells are heavy, expensive, lack impact robustness, and suffer from inertial noise during periods of high accelerations. The main reason these sensors are not suitable is because the sensor is part of the load path, and all the interaction forces from the ground to the robot have to travel through the sensor for the forces to be measured.

In order to withstand the forces in the load path, while still ensuring that there is just the right amount of deflection, the load cell structure has to be made out of heavy and stiff metals. If the structure deflects too much, then the strain gauge will deform plastically and become ruined. If the structure deflects too little, then the resolutions for the force sensing will suffer.

What is different about the force sensors we are making is that instead of directly capturing the load at the path, we sample the stress distribution and use that to deduce the forces that led to this stress distribution. A monolithic sensor integrated in a footpad, capable of measuring large magnitude normal and shear force sensing is developed to apply this principle. In legged locomotion, the capability to measure large forces is needed, as during running impact, forces can rise to around 2.4 times of the runner’s body weight.

Because the load path now travels through the entire cross section of the footpad, we can make use of lightweight, compliant rubber materials instead of stiff and heavy metals for the structure of the footpad. Sampling the stress distribution also allows us to deduce the shear forces using machine learning techniques. With this data, we can know the force profile of the foot at touchdown, and use that to accurately predict the subsequent trajectory of the robot.

Footpad Sensor Prototypes

Using new design principles and methodologies, we have developed a multi-axis, large force detecting foot sensor for legged robots. It is a monolithic composite structure that is composed of a piezoresistive sensor array PCB completely embedded in a protective polyurethane rubber layer. The composite architecture allows for compliance and traction during ground contact, while deformation alters the measured stress distribution. Using machine learning, we are able to map the local sampling of stress inside the polymeric footpad to forces in three axes with high accuracy. This footpad sensor is intended for use on the MIT Cheetah to provide a complete picture of the ground interaction forces that is a necessity in enabling high-speed and dynamic ground locomotion.

Multiple versions of the footpad sensor has been developed, and the current version has an onboard ARM microcontroller embedded directly in the footpad to process the data and communicate with both serial and RS-422 protocols.

Footpad v1

Footpad v2

Footpad v3

 


 

Force Sensing Smart Shoes of the Future

According to Statista, the global sports footwear market is projected to hit $87 billion by 2020. More than 20 companies in this new and growing segment of ‘smart shoes’, including established brands such as Samsung, Nike, Adidas, Puma, and Under Armor.

There is also ongoing work in developing the next generation of the wearables in the form of smart shoes that are capable of real-time in-situ measurement of force data. This work will involve integrating the low-cost, lightweight, multi-axis force sensor that have been developed for use on the MIT Cheetah robot. The ultimate goal is to use these force sensing shoes to help assist the elderly and disabled during walking and for fall prevention and mitigation. Athletes can also benefit from the data collected during training to better optimize their workouts.

Android App for Wireless Data Streaming

We are currently creating an Android app to stream the data from the force sensing smart shoes. As the force sensing smart shoes produces a large amount of data due to the high bandwidth (1kHz), this is an especially challenging task to receive this data, process it quickly, and graph it in a timely manner.

 

 

 

 

 

 

Thanks to SciChart Ltd (www.scichart.com) for sponsoring the developer licenses used to make the Android app. Without SciChart, the high performance realtime graphs would not be possible.