Probabilistic Online Prediction of Robot Actions Results based on Physics Simulation ”Functional Imagination“ Sebastian Rockel TAMS, University of Hamburg, Hamburg Nov. 30, 2015
Journal of Economic Perspectives—Volume 29, Number 3—Summer 2015—Pages 51–60 Is a Cambrian Explosion Coming for Robotics? Gill A. Pratt Eight Technical Drivers 1. Exponential growth in computing performance 2. Improvements in electromechanical design tools and numerically controlled manufacturing tools 3. Improvements in electrical energy storage 4. Improvements in electronics power efficiency 5. Exponential expansion of the availability and performance of local wireless digital communications 6. Exponential growth in the scale and performance of the Internet 7. Exponential growth of worldwide data storage 8. Exponential growth in global computation power 2
Journal of Economic Perspectives—Volume 29, Number 3—Summer 2015—Pages 51–60 Is a Cambrian Explosion Coming for Robotics? Gill A. Pratt Cloud Robotics: Big Idea #3: Learning from Imagination 3
Outline 1. Related Work 2. Architecture & Integration 3. Uncertainty 4. Experiments & Results 5. Conclusion 4
Task Planning symbolism Table Table Table discreteness x,y,z … x,y,z x,y,z x,y,z x,y,z x,y,z continuity 5
Functional Imagination • improve task-planning based system • use prediction from physical simulation • integrate prediction • use predicted results to adapt plan execution • forestall failures 6
Related Work • concurrent reactive planning (Beetz 2000) • meta-CSP: hybrid planning/reasoning (Moffit et al. 2006) • functional imagination (Marques et al. 2008) • GTP + HTN (de Silva et al. 2013) 7
(2012-2015) • ”Robustness by Autonomous Competence Enhancement“ • high-level world representation • multi-level experience representation • learning and generalizing from experiences 8
RACE Architecture plan ROS OWL concepts actions Semantic ROBOT sensor HTN Execution data Planner OWL Monitor action Capabilities Ontology results initial state, plan, fluents, goal goal schedule new concepts plan continuous Conceptualizer ex − data periences Blackboard Perception fluents fluents OWL concepts fluents Perceptual ex − Reasoning fluents instructions periences Memory and Interpretation Experience instructions User Extractor/ Annotator Interface 9
Methodology hierarchical Sense Plan Act Sense Plan Imagine Act 10
Methodology (cont’d) reactive Plan Sense Act Plan Sense Act Imagine 11
Processing plan generation execution evaluation is yes no imagine action? projection simulation robot 12
Architecture Robot plan Capabilities Planner ROS actions Executor fluents imagine results actions fluents Black- Functional Simulation 1 Simulation 1 Simulation 1 board Imagination imagine actions 13
Architecture (cont’d) Imagination Client 1 (PC 1) Functional Imagination Client 1 (PC 1) Imagination Client 1 (PC 1) Imagination Executor Blackboard Black- Executor Blackboard Executor board Evaluation Translation Simulation 1 Simulation 1 Simulation 1 14
Plan Adaptation time execution act1 … … act<..> par1 imagination par2 par3 15
Plan Adaptation (cont’d) SHOP2: grasp_object_w_arm mug1 rightarm1 assume_manipulation_pose !pick_up_object leave_manipulation_pose manipulationareaeastcounter1 mug1 manipulationareaeastcounter1 rightarm1 rightarm1 assume_manipulation_pose !move_base_blind !move_torso manipulationareaeastcounter1 premanipulationareaeastcounter1 torsoupposture rightarm1 (S. Stock, Univ. Osnabrück) !imagine ?task ?arg1 ?arg2 !imagine !move_base_param ?area slow/fast 16
Execution & Imagination • projection : generate world state foreach !imagine action • translation : converting symbolic (blackboard) values in discrete coordinates • FI evaluates best confidence and shortest time • FI returns ordered list of parametrization + duration 17
Sampling • sampling: f s = 20 Hz • discretization: • carry tall object: z 18
Handling Uncertainty • variables: (event 1), (event 2), .. a 1 a 2 • empirical coefficients: (event 1), (event 2), .. c 2 c 1 • confidence: • carry tall object: 19
Assumptions & Limitations • discrete thresholds (empirically defined) • discrete (action) parameter set • one action per imagination (plan-step sync) • performance correlation to simulation granularity • exogenous events rarely occur • ”challenging“ simulation 20
Experiments 1. simulation validation 2. recognition & manipulation 3. serve a coffee 4. carry a tall object 21
Simulation Validation PMA NorthTable1 nearStartArea1 North table1 table2 PMA WestTable2 PMA EastTable2 PMA EastCounter1 PMA SouthTable1 counter1 22
Simulation Validation (cont’d) 30 25 Duration Mean in [s] 20 15 10 5 0 1 2 3 4 5 6 7 8 9 10 11 12 1 mb_f EC1/ET2 5 Standard Deviation in [s] 2 mb_s EC1/ET2 Simulation 3 mb_f EC1/NT1 PR2 4 4 mb_s EC1/NT1 5 mb_f ET2/EC1 3 6 mb_s ET2/EC1 7 mb_f NT1/EC1 2 8 mb_s NT1/EC1 1 9 mt U/D 10 mt D/U 0 11 ta U/T 1 2 3 4 5 6 7 8 9 10 11 12 12 ta T/U 23
Recognition & Manipulation Recognition Probability torso up torso down 0,58 0,56 0,54 0,51 0,49 front back left right 24
Serve a Coffee (ICRA 2014) PMA North NorthTable1 table1 counter PMA Fail SouthTable1 robot execution re-planning/imagination standard prediction 0 75 150 225 300 25
Carry a Tall Object (IROS 2015) 26
Carry a Tall Object (IROS 2015) slow fast 27
My code online available.. • https://github.com/buzzer/pr2_imagination • https://github.com/buzzer/tams_pr2 28
Conclusion • improved task-planning system with ”functional imagination“ • ”common-sense“ physics-based prediction (cf. Marques, De Silva) • enabling hybrid reasoning (cf. Moffit) • based on action parametrization (cf. Beetz) • probabilistic projection (sampling, confidence) • validation of simulation • new system on a PR2 (out-of-the-box HTN + Gazebo) 29
Future Research • improve the robot’s performance in changing and partly unknown worlds • partial plan imagination • continuous representations (parameter + sampling) • exploit temporal aspects (re-scheduling) • reactive perception + imagination • cloud robotics 30
Thank you for your attention! 31
References Lavindra de Silva, Amit Kumar Pandey, and Rachid Alami. An interface for interleaved symbolic-geometric planning and backtracking. In Proceedings of the International Conference on Intelligent Robots and Systems (IROS) , 2013. Hugo Gravato Marques, Owen Holland, and Richard Newcombe. A modelling framework for functional imagination. In AISB Convention , pp. 51–58, 2008. Michael D. Moffitt and Martha E. Pollack. Optimal rectangle packing: A meta- CSP approach. In ICAPS, pp. 93-102. AAAI, 2006 Michael Beetz. Concurrent Reactive Plans: Anticipating and Forestalling Execution Failures. Springer-Verlag, Berlin, Heidelberg, 2000. Vere, S. A. Planning in Time: Windows and Durations for Activities and Goals. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI- 5(3): 246-247, 1983. 32
Recommend
More recommend
