Reinforcement Learning for Grasping Evaluation
Our focus is on the tradeoff between sensing cost and performance. When designing tactile sensors, the challenge lies in balancing the tradeoff between costly, highly accurate sensing and affordable alternatives with less information. In our study [3], we revaluated this tradeoff via Reinforcement Learning (RL). We trained the model with sensor-rich rewards while providing varied tactile sensing levels in the state vectors in RL. Surprisingly, our results show that, even with no tactile sensing in the state vectors, if the model is trained with rewards containing full tactile information, it performs on par with when trained with full tactile information in the state vectors. This challenges the conventional trade-off, suggesting that it’s possible to train models on tactile-enabled robot hands and deploy them to robot hands with minimal sensing. This streamlined workflow significantly cuts costs while maintaining high performance, making robotic grasping accessible beyond lab and warehouse settings and into the daily lives of communities.
Researchers: Zeo Liu
Low-Cost Tactile Sensors for
Robot Hands
Our focus is on affordable tactile sensor development, meeting the requirements for reliable grasping. While highly instrumented robot hands excel in controlled lab settings, household integration demands affordability and resilience to unexpected collisions. Current approaches, employing technologies like computer vision and costly tactile sensing, face challenges in occlusion, cost, consistency, interpretability, and accuracy. We developed a cost-effective solution, featuring a novel tactile sensor and a machine-learning pipeline for localizing and quantifying disturbance forces when robots encounter obstacles. This design spans a wide contact range and is sensitive to forces in three axes. The resulting machine learning model accurately estimates disturbance location and force on various unseen everyday objects, enabling the robot to plan responses and recover from grasp errors for reliable performance.
Researchers: Zeo Liu
Tactile Reinforcement Learning for Grasping with Low-Cost Deployment
Robotic hands equipped with tactile sensors are revolutionizing fields like medical robotics, rehabilitation, and industrial automation by enabling high-precision tasks, particularly in grasping. Our project focuses on developing Deep Reinforcement Learning models that leverage data from tactile sensors to train robotic hands in grasping strategies for unconstrained environments. Recognizing the cost barriers associated with sophisticated tactile sensors, we propose a cost-effective method that employs high-quality sensors during the training phase only. This approach aims to enhance the learning quality of the grasping models, which can then be deployed using more economical sensors without compromising performance.
Researchers: Ludovico Papavassiliou and Zeo Liu
Dexterous Manipulation
Robotic hands are important for tasks in many domains, from warehouses to disaster zones to people’s homes. However, current industrial grippers lack dexterity and generality, and anthropomorphic research hands are expensive, fragile, and difficult to control. A task-centric design methodology holds considerable promise to deliver inexpensive, robust hands.
The i-HY hand was developed under DARPA’s Autonomous Robotic Manipulation – Hardware Track (ARM-H) program with the goal of creating durable, inexpensive robotic hands capable of manipulation as well as grasping, suitable for use on mobile robots The novel minimalistic design of i-HY was developed by choosing a set of target tasks around which the design of the hand was optimized. This resulted in an underactuated hand that is capable of performing a wide range of grasping and in-hand manipulation tasks using only 5 motors. To develop this, the principles driving underactuated grasping were extended to in-hand manipulation, and fingers were designed that are capable of both firm power grasps and low-stiffness fingertip grasps using only the passive mechanics of the finger mechanism. Experimental results demonstrate successful grasping of a wide range of target objects, the stability of fingertip grasping, as well as the ability to adjust the force exerted on grasped objects using the passive finger mechanics. The project was a collaboration between iRobot, Harvard University, and Yale University.
Researchers: Leif Jentoft, Yaroslav Tenzer, and Lael U. Odhner
Video
Numerical Optimization of
Robotic Hands
The design of robotic hands for dexterous grasping and manipulation often leads to fully-actuated, anthropomorphic (biomimetic) solutions. These solutions, however elegant, typically entail complex actuation frameworks that make them expensive and difficult to implement effectively. Underactuated, compliant robotic hands exploit passive mechanics and joint coupling to reduce the number of actuators required to achieve dexterous, robust grasping in unstructured environments. While reduced actuation requirements generally serve to decrease design cost and improve grasp planning efficiency, overzealous simplification of an actuation topology, coupled with insufficient tuning of mechanical compliance and hand kinematics, can adversely affect grasp quality and adaptability. In this research, we develop a computational framework for systematically reducing the complexity and cost of robotic hands while promoting improved hand dexterity and grasp robustness.
Researchers: Frank Hammond III
Intelligent Passive Mechanics
SDM Hand
Compliance conveys several advantages for robotic grasping. In unstructured environments, sensing uncertainties are large, and the target object size and location may be poorly known. Finger compliance allows the gripper to conform to a wide range of objects while minimizing contact forces. Robot joint compliance or stiffness has often been considered in the context of active control, where active control uses sensors and actuators to achieve a desired force-deflection relationship. In contrast, passive compliance, implemented through springs in robot joints, offers additional benefits, particularly in impacts, where control loop delays may lead to poor control of contact forces. The reduced need for sensing required to create active compliance can also lead to lower implementation costs.
Researchers: Aaron Dollar and Leif Jentof
Video
Compliance and Sensing
Compliant joints provide a number of advantages for grasping, but they also open new approaches to sensing. They enable gentle interactions even with low-bandwidth, position-controlled hands and have few blind spots. We have developed new sensors to measure the deflection of compliant flexure joints, and algorithms to use them to determine object geometry and contact forces.
Object geometry is important for grasp planning and object classification. However, existing approaches for tactile sensors are slow and require expensive or delicate hardware. Joint-angle sensors can be used to determine the surface geometry of a target object when used in conjunction with compliant joints. This approach is mechanically robust, inexpensive, and requires only basic control of manipulator position (i.e., no force or impedance control).
Researchers: Leif P. Jentoff
Intelligent Passive Mechanics
SDM Hand
Although the human hand has a very capable sensor suite, thirty years of benchtop research have failed to deliver tactile sensing suites that are robust and inexpensive enough to provide grasping information at a practical cost/benefit ratio. We are developing hardware that is inexpensive, robust, and matched to the capabilities of robotic rather than biological fabrication techniques, and perception algorithms to process this data that are robust to the noise and uncertainty present in unstructured environments.
Researchers: Aaron Dollar and Leif Jentof
Video
Robust Robotic Mechanisms and Sensors via Shape Deposition Manufacturing
One of the greatest successes of biologically inspired design has been the development of mechanically robust robots. One promising biomimetic fabrication technique is Shape Deposition Manufacturing (SDM), which alternates material deposition and machining to produce robot structures with compliant joints and embedded sensing and actuation elements. We explore the benefits of using Shape Deposition Manufacturing for constructing a simple two-fingered gripper and add to the tools available to robot designers by developing a range of sensing modalities compatible with the process. These include Hall-effect sensors for joint angle sensing, embedded strain gauges for 3-axis force measurements, optical reflectance sensors for tactile sensing, and piezoelectric polymers for contact detection. In addition to a simple construction process, the resulting parts are extremely robust, fully functional after high-impact loads and other forces due to unintended contact.
Researchers: Aaron M. Dollar
