Work

High precision grasp pose detection in dense clutter

Year: 2016
Type: conference-paper
Abstract: This paper considers the problem of grasp pose detection in point clouds. We follow a general algorithmic structure that first generates a large set of 6-DOF grasp candidates and then classifies each of them as a good or a bad grasp. Our focus in this paper is on improving the second step by using depth sensor scans from large online datasets to train a convolutional neural network. We propose two new representations of grasp candidates, and we quantify the effect of using prior knowledge of two forms: instance or category knowledge of the object to be grasped, and pretraining the network on simulated depth data obtained from idealized CAD models. Our analysis shows that a more informative grasp candidate representation as well as pretraining and prior knowledge significantly improve grasp detection. We evaluate our approach on a Baxter Research Robot and demonstrate an average grasp success rate of 93% in dense clutter. This is a 20% improvement compared to our prior work.
Language: English
Open Access status: green
Awards: THIS PROJECT IS AIMED AT ADVANCING CAPABILITIES OF NASAS HUMANOID ROBOT R5 THROUGH THE DEVELOPMENT OF NOVEL HIGH LEVEL BEHAVIORS PERCEPTION BASED MANIPULATION AND GRASPING AND OPERATOR INTERFACES. WE PROPOSE TO HOST ONE OF R5S AT THE NEW ENGLAND ROBOTICS VALIDATION AND EXPERIMENTATION (NERVE) CENTER AT UMASS LOWELL IN LOWELL MASSACHUSETTS LESS THAN 30 MILES FROM BOSTON. AT 10 000 SQUARE FEET THE NERVE CENTER IS THE LARGEST INDOOR ROBOT TEST FACILITY IN NEW ENGLAND. THE COMPLEMENTARY EXPERTISE OF THE PROPOSING RESEARCH TEAM IN MODEL BASED ROBOT CONTROL GRASPING AND HUMAN ROBOT INTERACTION WILL BE UNIFIED TO DEVELOP AUTONOMOUS BEHAVIORS FOR R5. THE TEAMS COLLECTIVE EXPERIENCE FROM THE DARPA ROBOTICS CHALLENGE (DRC) FINALS WILL ENABLE THE DEVELOPMENT OF NEW CAPABILITIES FOR R5 TO EXECUTE A SET OF EXPECTED NASA ADMINISTERED SPACE ROBOTICS CHALLENGE (SRC) TASKS. THE DEVELOPMENT EFFORT WILL BE CARRIED OUT IN THREE MAJOR THRUSTS ON CONSTRAINED MOTION PLANNING AND CONTROL PERCEPTION MANIPULATION AND GRASPING AND USER INTERFACE DESIGN. IN COMPLETING SPACE EXPLORATION RELEVANT MANIPULATION AND MOBILITY TASKS FOR THE SRC AN ITERATIVE MODELBASED DESIGN APPROACH WILL BE USED TO SPECIFY AND VERIFY REQUIREMENTS FOR EACH TASK DEVELOP MODELS FOR PHYSICAL HUMAN AND ROBOT ACTIONS AS WELL AS THE ENVIRONMENT SELECT AND COMPOSE MODELS OF COMPUTATION SIMULATE THE HUMAN SUPERVISED SYSTEM AS A WHOLE AND VERIFY AND VALIDATE THE ALGORITHM DESIGNS ON THE PHYSICAL ROBOT. THE HIGH LEVEL APPROACH TO TASK DEVELOPMENT AND VALIDATION IS TO START WITH TELEOPERATION AND THEN TO MOVE TOWARD PARTIAL OR FULL AUTONOMY THROUGH THE INTEGRATION OF NEW CAPABILITIES. THE RESEARCH TEAM WILL REPURPOSE MOTION PLANNING FULL BODY CONTROL PERCEPTION MANIPULATION GRASPING AND OPERATOR INTERFACES CAPABILITIES DEVELOPED WITHIN THE SCOPE OF RECENT DRC FINALS TO SYSTEMATICALLY VALIDATE SRC TASKS. IN COLLABORATION WITH NASA ENGINEERS THE LESSONS LEARNED FROM THIS VALIDATION EFFORT WILL BE USED TO REFINE THE SRC TASKS. IN YEAR 2 OF THE PROJECT WE WILL HOST AN SRC COMPETITOR AT THE NERVE CENTER AND PROVIDE THE COMPETITOR TEAM WITH ACCESS FOR TESTING THEIR ALGORITHMS ON R5. UPON COMPLETION THIS PROJECT WILL ADVANCE THE TECHNOLOGICAL READINESS OF HUMANOID ROBOT AUTONOMY FOR SUCCESS IN FUTURE SPACE MISSIONS., RI: Large: Collaborative Research: Reconstructive recognition: Uniting statistical scene understanding and physics-based visual reasoning, MANIPULATING SMALL FLEXIBLE AND DEFORMABLE OBJECTS IS A KEY CHALLENGE IN DEVELOPING ROBOTS TO WORK WITH HUMANS IN MANUFACTURING AND SPACE APPLICATIONS. IN MANUFACTURING DOMAINS ASSEMBLY TASKS OFTEN INVOLVE SMALL OR FLEXIBLE MATERIALS SUCH AS CABLES FABRIC CARPETING FLEXIBLE PLASTIC ETC. IN NASA EXTRA-VEHICULAR REPAIR TASKS IT IS OFTEN NECESSARY TO MANIPULATE THERMAL BLANKETS IN ORDER TO ACCESS EQUIPMENT. WE PROPOSE USING FORCE AND TACTILE SENSING TO LOCALIZE THESE KINDS OF MATERIALS ACCURATELY DURING IN-HAND MANIPULATION AND THEREBY MAKE DEXTEROUS ROBOT MANIPULATION MORE ROBUST AND PRECISE. PRIOR WORK BY THE PI USING THE NASA-JSC ROBONAUT 2 HAS DEMONSTRATED THAT TACTILE SENSING CAN INDEED BE USED TO MAKE MANIPULATION MORE ROBUST. HOWEVER IT IS DIFFICULTTO SCALE THE APPROACH UP TO LARGE PROBLEMS BECAUSE OF THE CHALLENGES INVOLVED IN DEVELOPING A HAPTIC MODEL OF WHAT THE MATERIAL IS EXPECTED TO FEEL LIKE. THE PROBLEM IS THAT IT IS HARD TO MODEL FLEXIBLE MATERIALS ANALYTICALLY AND IT CAN BE DIFFICULT TO GATHER A SUFFICIENTLY LARGE AMOUNT OF TRAINING EXPERIENCE TO TRAIN AN ACCURATE MODEL. MOREOVER EVEN VERY SOPHISTICATED TACTILE SENSORS SOMETIMES PRODUCE VERY NOISY AND ATTENUATED DATA.THIS PROPOSAL IS TO MAKE FLEXIBLE MATERIALS MANIPULATION MORE PRACTICAL BY DEVELOPING NEW METHODS OF CREATING AND USING HAPTIC MODELS. WE PROPOSE TO LEVERAGE EXISTING ROBOT LOCALIZATION AND MAPPING WORK DEVELOPED IN THE MOBILE ROBOT COMMUNITY. ALREADY IN OUR LAB WE HAVE STARTED THIS WORK AND DEVELOPED AND DEMONSTRATED SOME BASIC TACTILE MAPPING TECHNIQUES THAT WERE SUMMARIZED AT THE 2013 NASA NRI PI MEETING. HERE WE PROPOSE TO APPLY THIS WORK IN REAL ROBOT MANIPULATION CONTEXTS CULMINATING IN A DEMONSTRATION OF TACTILE MAPPING AND LOCALIZATION CAPABILITIES IN THE CONTEXT OF SMALL AND FLEXIBLE MATERIALS MANIPULATION TASKS. IN YEAR 1 WE PROPOSE TO CONTINUE TO EXPLORE TACTILE REGISTRATION AND MAPPING TECHNIQUES AND TO DEMONSTRATE BASIC TACTILELOCALIZATION CAPABILITIES IN THE CONTEXT OF A FINE-MATERIALS INSERTION TASK. WE EXPECT THIS WORK TO BE DEMONSTRATED IN THE CONTEXT OF AN EXPERIMENT WITH THE BAXTER ROBOT LOCATED AT NORTHEASTERN UNIVERSITY. IN YEAR 2 WE PROPOSE TO EXPAND THESE MANIPULATION CAPABILITIES BY DEVELOPING MORE ROBUST AND SOPHISTICATED PLANNING AND CONTROL METHODS THAT REASON ABOUT HOW THE MANIPULATION STRATEGY AFFECTS THE INFORMATION CONTENT AND SAFETY ENVELOPE OF THE MANIPULATION TASK., NRI: Collaborative Research: Human-Supervised Perception and Grasping in Clutter
OpenAlex ID: W2290564286
FWCI: 48.76
Cites: 21
Cited by: 359
Related to: 10