Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Soft-Robotic Learning for Industrial Assembly (IJCAI'20 WS Invited Talk)

Soft-Robotic Learning for Industrial Assembly (IJCAI'20 WS Invited Talk)

Masashi Hamaya and Yoshihisa Ijiri, "Soft robotic learning for industrial assembly", presented @ Knowledge Based Reinforcement Learning Workshop, IJCAI'20.

753eb5a167cf43a033413f08e63f3632?s=128

OMRON SINIC X

March 30, 2021
Tweet

Transcript

  1. © 2021 OMRON SINIC X Corporation. All Rights Reserved. Soft

    Robotic Learning for Industrial Assembly (Invited Talk) Yoshihisa Ijiri and Masashi Hamaya (OMRON SINIC X) Invited talk at Knowledge Based Reinforcement Learning Workshop in IJCAI 2020 Jan. 7, 2021
  2. Soft Robotic Learning for Industrial Assembly Yoshihisa Ijiri and Masashi

    Hamaya OMRON SINIC X Corporation
  3. 2 © 2020-2021 OMRON Corporation All Rights Reserved What is

    OMRON Group doing? Switches relay, and MEMS sensors for consumer electronics 300 thousands+ lineups of industrial sensors, controllers and robots, leading the innovation of global manufacturing 5% 18% 13% 13% 51% Sense & control + think Core technologies Others Business under the Direct Control of HQ Digital Healthcare Social Infrastructures Solution & Service Electronic & Mechanical Components Factory Automation Surveillance, marketing, railway, highway solutions Blood pressure meter, thermometer, etc. PV inverters, Storage systems, etc.
  4. 3 © 2020-2021 OMRON Corporation All Rights Reserved Needs for

    industrial robotics https://morunuma.keymary.net/robotfactory/ https://www.khi.co.jp/rd/magazine/163/nj163tr01.html Product mix Tailored production Mass production High-mix low-volume production Quantity Automated Human
  5. 4 © 2020-2021 OMRON Corporation All Rights Reserved Launching robotics

    system in industrial lines involves a huge amount of engineering, which makes it difficult to automate high-mix low-volume productions. Huge amount of engineering… Sensing developments؟~1 month Robotic system developments؟2~3 months Sensing Robots & jigs Task development & optimization؟2~3 weeks System re-design & re-optimization: 1 month Trouble Launch Operation
  6. 5 © 2020-2021 OMRON Corporation All Rights Reserved What makes

    it so difficult? Product assembly is always a combination of contact-rich tasks Connector insertion Circuit board insertion Fitting of upper and lower housings Screwing Snap-lock insertion ؞ ؞ ؞ Even a very small error is critical for engagement of two parts https://www.tsukumo.co.jp/bto/faq/120005.html https://news.livedoor.com/article/detail/13932272/
  7. 6 © 2020-2021 OMRON Corporation All Rights Reserved What makes

    it so difficult? Product assembly is Always about manipulation of diverse parts types of screws types of packages types of connectors Different sizes, shapes, frictions, … ؞ ؞ ؞ That's one small thing for a man, one giant leap for a robot.
  8. 7 © 2020-2021 OMRON Corporation All Rights Reserved So in

    current robotics Jigs and careful calibration for error elimination Careful motion programming for preventing catastrophic failure Elaborated hands and jigs for diverse parts These make current robotics costly and complicated It is too difficult for ordinary people to get advantage of robots
  9. 8 © 2020-2021 OMRON Corporation All Rights Reserved Assembly is

    still very challenging because of geometric uncertainty! Assembly without jigs and special hands
  10. 9 © 2020-2021 OMRON Corporation All Rights Reserved Example of

    simple assembly From the WRC2018 assembly task Our arms and hands are not as precise as robot ones but we can do it very easily
  11. 10 © 2020-2021 OMRON Corporation All Rights Reserved Observation: how

    can we do it so easily? From the WRC2018 assembly task Softness? To avoid catastrophic failure, and absorbing errors
  12. 11 © 2020-2021 OMRON Corporation All Rights Reserved Core idea

    Uncertainty-tolerant robotics by Softness × Learning × Tactile Learning Softness Vision/ Tactile Safety Compensate unpredictability Endurance & active touch Nonlinear Modeling Efficient representation Richer information Tactile & force sensing Soft Robotics Data-driven control
  13. 12 © 2020-2021 OMRON Corporation All Rights Reserved Developments for

    soft robot hardware and learning control methods 1. Hardware development and sample efficient learning control 3. Soft robots with knowledge of humans 2. Soft robots with knowledge of previous experiences • Activable soft wrist • Sample efficient model-based RL • Transfer RL • Sim-to-real transfer • Learning from demonstrations
  14. 13 © 2020-2021 OMRON Corporation All Rights Reserved • Soft

    robots can handle contact-rich tasks. • Cable-driven soft wrist allows changing between hard and soft mode. A compact, cable-driven, activatable soft robot wrist for assembly [Drigalski and Tanaka et al., IROS 2020] Hardware development Peg-in-hole task Soft and rigid mode Design of soft wrist Soft wrist No soft wrist
  15. 14 © 2020-2021 OMRON Corporation All Rights Reserved • Modeling

    difficulty due to softness and contact richness • Task segmentation and model-based reinforcement learning [Deisenroth et al., 2015] • Robot leaned each subtask with a few trials. Learning Robotic Assembly Tasks with Lower Dimensional Systems by Leveraging Physical Softness and Environmental Constraints [Hamaya et al., ICRA 2020] Sample efficient soft robotic assembly learning Learning “fit” subtask After learning
  16. 15 © 2020-2021 OMRON Corporation All Rights Reserved • Transfer

    reinforcement learning is promising for adaptation to unknown tasks. • Communication with source environments is sometimes difficult. MULTIPOLAR: Multi-Source Policy Aggregation for Transfer Reinforcement Learning between Diverse Environmental Dynamics [Barekatain et al., IJCAI 2020] Learning from knowledge of previous experiences Peg-in-hole tasks with different pegs and holes
  17. 16 © 2020-2021 OMRON Corporation All Rights Reserved • MULTIPOLAR:

    MULTI-source POLicy AggRegation • Able to work on black-box source policies Learning from knowledge of previous experiences State ݏ௧ Auxiliary network for predicting residuals: ܨୟ୳୶ ݏ௧ ; ߠୟ୳୶ ߠaux Continuous action space: ߨ୲ୟ୰୥ୣ୲ ؠ ࣨ ܨ ݏ௧ ; ܮ, ߠୟ୥୥ , ߠୟ୳୶ , ȭ Discrete action space: ߨ୲ୟ୰୥ୣ୲ ؠ ܵ݋݂ݐܯܽݔ ܨ ݏ௧ ; ܮ, ߠୟ୥୥ , ߠୟ୳୶ ߤଵ ߤଶ ߤ௄ … Source policies ܮ = ߤଵ , … , ߤ௄ … ٖ ܣ௧ … ߠagg Adaptive aggregation of source policies: ܨୟ୥୥ ݏ௧ ; ܮ, ߠୟ୥୥ ܨ ݏ௧ ; ܮ, ߠୟ୥୥ , ߠୟ୳୶ + MULTIPOLAR: Multi-Source Policy Aggregation for Transfer Reinforcement Learning between Diverse Environmental Dynamics [Barekatain et al., IJCAI 2020]
  18. 17 © 2020-2021 OMRON Corporation All Rights Reserved • MULTIPOLAR

    showed higher performances than learning from scratch. Learning from knowledge of previous experiences MULTIPOLAR Learning from scratch Ant Hopper MULTIPOLAR: Multi-Source Policy Aggregation for Transfer Reinforcement Learning between Diverse Environmental Dynamics [Barekatain et al., IJCAI 2020] Our session will start at Jan 13th 18:00-19:20 (JST)!
  19. 18 © 2020-2021 OMRON Corporation All Rights Reserved • Further

    sample efficiency for real robot transfer by aggregating dynamics models • Employ model-based RL [Chua et al., 2018] TRANS-AM: Transfer Learning by Aggregating Dynamics Models for Soft Robotic Assembly [Tanaka et al., under review] Learning from knowledge of previous experiences Dynamics model aggregation Transfer to different hole angles
  20. 19 © 2020-2021 OMRON Corporation All Rights Reserved • Sim-to-real

    transfer can also reduce engineering efforts. • EXI-Net can for multiple environments. • Conditioned with explicit (e.g., mass, friction) and implicit parameters (e.g., object shapes) • Online adaptation with parameter estimation • It can be also applied to soft robots. EXI-Net: EXplicitly/Implicitly Conditioned Network for Multiple Environment Sim-to-Real Transfer [Murooka et al., CoRL 2020] Learning from knowledge of previous experiences Training in simulation Test in real world Sim-to-real transfer in multiple object pushing tasks Simulated soft robot
  21. 20 © 2020-2021 OMRON Corporation All Rights Reserved • Designing

    reward function is difficult. • Exploit successful and failed demonstrations • Develop a teaching device that mimics the robot’s gripper Learning Soft Robotic Assembly Strategies from Successful and Failed Demonstrations [Hamaya et al., IROS 2020] Learning from knowledge of humans Teaching device Learning given demonstrations
  22. 21 © 2020-2021 OMRON Corporation All Rights Reserved • Soft

    robotics for industrial assembly • Uncertainty-tolerance thank to softness • Our research examples with soft robots • Cable-driven soft wrist to avoid unnecessary oscillation • Sample efficient RL based control to tackle modeling difficulty • Learning from knowledge of previous experiments using transfer RL and sim-to-real • Learning from knowledge of humans using human demonstrations Summary
  23. 22 © 2020-2021 OMRON Corporation All Rights Reserved • Learning

    with tactile sensors • In-hand-pose estimation [Drigalski et al., ICRA 2020] • In-hand manipulation for blind bin picking [Ishige et al., IROS 2020] • Combining assembly strategies with tactile sensors • Safe and recovery RL with physical and computational safety Future works In-hand pose estimation [Drigalski et al., ICRA 2020] In-hand manipulation [Ishige et al., IROS 2020]
  24. 23 © 2020-2021 OMRON Corporation All Rights Reserved Shameless announcement

    Diabolo control [Drigalski and Joshi et al., under review] We are hiring internship students and full researchers. If you are interested in positions at OMRON SINIC X, please email us. internships@sinicx.com
  25. None