Toward data-driven contact mechanics

Size: px

Start display at page:

Download "Toward data-driven contact mechanics"

Morgan Warner
5 years ago
Views:

1 Toward data-driven contact mechanics KRIS HAUSER Shiquan Wang and Mark Cutkosky Joint work with: Alessio Rocchi NSF NRI #

2 Analytical contact models are awesome Side Force closure but not form closure Not force closure Top

4 DRC-Sim in the DARPA Robotics Challenge

5 DRC-Hubo Planners and controllers tested in simulation before lab testing

6 Contact mechanics is not just Coulomb models But most of our tools are based on Coulomb (and other relatively simple) models Wait for a phenomenon to be modeled then incorporated to our tools?

7 When sims go wrong

8 Gazebo Open Dynamics Engine, OPCODE collision module DRCSim, Gazebo (video from IHMC)

9 Grasping with compliant hands Obvious physically implausible artifacts

10 Simulations are getting better Match physical and simulation protocol Grasp, lift, shake, open hand (Rocchi et al, Stable simulation of underactuated compliant hands, ICRA 2016)

11 What aren t we modeling? Articulation, pages Deformation Bristles, nonuniform friction (Plastic packaging, flexing, )

12 Amazon Picking Challenge Aerodynamics of a vacuum gripper in partial contact with a deformable body? 12

13 Compliant robot grippers Tendons, flexible surface coverings, flexible joints, gearing, pulley friction Righthand Robotics ReFlex IIT/Pisa SoftHand

14 Exotic grippers Adhesion / anisotropic forces useful for locomotion, manipulation Suction Electrostatic Spines Van der Waals Granular jamming RiSE LEMURIIb Spinybot Stickybot CMU/MPI EPFL Cornell/Chicago

15 What do we want from our tools? Fidelity to real world Out-of-sample generalization Facilitate human engineer s creativity Speed (maybe)

16 What to learn? Partial dynamic information Parameters for physics sims Parameters for mathematical models Dynamics Dynamics function Rigid body simulations Partial decisions Simplified grasps Grasp poses Target corrections Decisions Policies Force limit surfaces Grasp metrics Tracking controllers Trajectories Pressure distributions Optimization margins

17 What to learn? Partial dynamic information Parameters for physics sims Parameters for mathematical models Force limit surfaces Pressure distributions Dynamics Dynamics function Rigid body simulations Grasp metrics Partial decisions Simplified grasps Grasp poses Target corrections Tracking controllers Optimization margins Decisions Policies Rely on optimization/ planning to achieve manipulation Trajectories

18 What to learn? Partial dynamic information Parameters for physics sims Parameters for mathematical models Dynamics Zhou and Hauser, RSS 2017 Revisiting Contact workshop Learn control directly Force limit surfaces Dynamics function Rigid body simulations (May need a perception system, or might learn endto-end) Pressure distributions Grasp metrics Partial decisions Simplified grasps Grasp poses Target corrections Tracking controllers Optimization margins Decisions Policies Trajectories

19 What to learn? Partial dynamic information Parameters for physics sims Parameters for mathematical models Dynamics Dynamics function Rigid body simulations Partial decisions Simplified grasps Grasp poses Target corrections Decisions Policies Learning full dynamics / decisions Tracking requires less analysis from a human controllers engineer Force limit surfaces Pressure distributions Grasp metrics Optimization margins Typically requires a lot of data Trajectories

20 What to learn? Partial dynamic information Parameters for physics sims Parameters for mathematical models Dynamics Dynamics function Rigid body simulations Partial decisions Simplified grasps Grasp poses Target corrections Decisions Policies Partial learning combines data with Tracking prior knowledge of physics controllers Force limit surfaces Pressure distributions Grasp metrics Optimization margins Generalizes better with less data Trajectories

Parameters for mathematical models Force limit surfaces Pressure distributions Dynamics Dynamics function Rigid body simulations

21 What to learn? Hauser, Wang, and Cutkosky, RSS 2017 Luo and Hauser, RSS 2015 & AuRo 2017 Partial dynamic information Parameters for physics sims Parameters for mathematical models Force limit surfaces Pressure distributions Dynamics Dynamics function Rigid body simulations Grasp metrics Partial decisions Simplified grasps Grasp poses Target corrections Partial learning combines data with Tracking prior knowledge of physics controllers Optimization margins Generalizes better with less data Decisions Policies Trajectories

22 3D Geometry in Simulation

23 3D Geometry in Simulation IROS 2016 Robot Grasping and Manipulation Challenge simulation track YCB Object Set (above), APC 2015 dataset, Princeton Shape Benchmark

with Robosimian Microspine unit: spring loaded needles latch onto

24 Microspine hands for vertical climbing [Hauser, Wang, Cutkosky, RSS 2017] Versatile Locomotion project: toward human-scale rock climbing with Robosimian Microspine unit: spring loaded needles latch onto asperities in rock Large downward shear, adhesion per unit Unit Palm Hand Vision of robot

25 Sloper Crimp

26 Admissible force volume modeling Define contact patch and local frame Measure max forces in all directions to obtain admissible volume Stochastic single-needle aggregation model agrees well with experimental data Shear adhesion Large shear in +x axis Lateral shear adhesion

j, then A j f b j is satisfied for A j, b j defined by

27 Admissible force volume modeling Perform convex decomposition of admissible volume If force is in region j, then A j f b j is satisfied for A j, b j defined by halfplanes Feasible force iff A 1 f b 1 A c f b c Convex decomposition

MILP equilibrium testing Given contact patch reference frames and external load w ext Formulate wrench matrix W Define Boolean indicators z i,j for each region in convex decomposition Load MILP Find

28 MILP equilibrium testing Given contact patch reference frames and external load w ext Formulate wrench matrix W Define Boolean indicators z i,j for each region in convex decomposition Load MILP Find f, z 1,1 z k,c such that Wf + w ext = 0 A i,j f i b i,j + BIG 1 z i,j for i = 1,, n and j = 1,, c z i,j {0,1} for all i, j c z i,j = 1 for all i σ j=1 F i,1 F i,2 Bounding volume hierarchy + custom BnB solver heuristics: ~20-600ms for 8-48 CPs, orders of magnitude faster than SCIP, Gurobi H i

29 Validation on physical passive 2-finger gripper Wrench space calculations on SpinyHand Power grasp Good agreement between predictions and experiments (11/14 slip events < 5 ) Some correctly predicted slip events Parallel grasp Three incorrectly predicted slip events FP FN FN

Articulated robot tracking is hard Associating data in physical trials with analytically

30 Thoughts 3D geometry capture: How to get inertial parameters, contact parameters, deformation? Noise? Pai et al 2001 Interaction capture tools? Articulated robot tracking is hard Associating data in physical trials with analytically simulated elements is hard Matching simulated parameters (e.g., reference frames, internal structures) to world Identifying relevant features in data

31 Recap Toward generalization + fidelity by combining data + analytical physics Robotics needs better datadriven simulation tools Let s work together! Duke s Pratt School of Engineering is hiring tenure-track faculty in robotics Contact me for more information Thank you! NSF NRI #

Can we quantify the hardness of learning manipulation? Kris Hauser Department of Electrical and Computer Engineering Duke University

Can we quantify the hardness of learning manipulation? Kris Hauser Department of Electrical and Computer Engineering Duke University Robot Learning! Robot Learning! Google used 14 identical robots 800,000