sapien
play

SAPIEN A SimulAted Part-based Interactive ENvironment Fanbo Xiang 1 - PowerPoint PPT Presentation

SAPIEN A SimulAted Part-based Interactive ENvironment Fanbo Xiang 1 Yuzhe Qin 1 Kaichun Mo 2 Yikuan Xia 1 Hao Zhu 1 Fangchen Liu 1 Minghua Liu 1 Hanxiao Jiang 3 Yifu Yuan 5 He Wang 2 Li Yi 4 Angel X. Chang 3 Leonidas J. Guibas 2 Hao Su 1 1 UC San


  1. SAPIEN A SimulAted Part-based Interactive ENvironment Fanbo Xiang 1 Yuzhe Qin 1 Kaichun Mo 2 Yikuan Xia 1 Hao Zhu 1 Fangchen Liu 1 Minghua Liu 1 Hanxiao Jiang 3 Yifu Yuan 5 He Wang 2 Li Yi 4 Angel X. Chang 3 Leonidas J. Guibas 2 Hao Su 1 1 UC San Diego 2 Stanford University 3 Simon Fraser University 4 Google Research 5 UC Los Angeles

  2. About Me Fanbo Xiang, UCSD UIUC BS CS, BS Math UCSD: MS CS Advisor: Hao Su, UCSD Research: Graphics, Vision, Robotics

  3. Outline ● Background ○ Intelligence, vision and robotics ○ Simulated environment ● SAPIEN architecture ○ Physics and robotics ○ Renderer ○ Assets ● Future research problems

  4. Intelligent Agent

  5. Perceive information Intelligent Agent

  6. Perceive information Intelligent Agent Adapt behavior

  7. Perceive information Adapt behavior

  8. Perceive information Computer vision Adapt behavior

  9. Perceive information Computer vision Robotics control Adapt behavior

  10. Perceive information Computer vision Robotics control Adapt behavior

  11. Perceive information Computer vision Planning? Active perception? Robotics control Adapt behavior

  12. Data driven approaches

  13. Learn from data?

  14. Real-world data collection

  15. Reinforcement Learning? Manual data collection?

  16. Reinforcement Learning

  17. RL Agent

  18. Observation RL Agent

  19. Perception Observation RL Agent

  20. Perception Observation RL Agent Action

  21. Perception Observation RL Agent Action Control

  22. Perception Observation State update RL Agent Reward Action Control

  23. RL Problems ● Expensive hardware ● Slow data collection ● Low sample efficiency ● Overfit to specific agent

  24. Imitation Learning (Manual Data Collection)

  25. Perception Observation State update Human Action Control

  26. Perception Observation Training State update Human IL Agent Action Control

  27. IL Problems ● Unintuitive control ● May not be optimal ● Human expertise

  28. Simulated Environment

  29. Manual data collection RL ● More freedom in controller ● Will not break hardware Simulated Environment design ● Easy to scale ● Less expertise required

  30. Simulated Environment

  31. Simulated Environment ? Real World

  32. Simulated Environment Physics

  33. Simulated Environment Physics Robotics

  34. Simulated Environment Physics Robotics Rendering

  35. Simulated Environment Physics Robotics Rendering Simulation Content

  36. Physics Robotics Rendering Simulation Content SAPIEN

  37. SAPIEN SAPIEN SAPIEN Engine Renderer Asset Physics Robotics Rendering Simulation Content

  38. SAPIEN Engine

  39. SAPIEN Engine SAPIEN Engine PhysX Physical Simulator

  40. SAPIEN Engine SAPIEN Engine PhysX Physical Simulator Articulation World Interface Interface

  41. SAPIEN Engine SAPIEN Engine PhysX Physical Simulator Articulation World Interface Interface ROS Interface Controller Sensor Interface Interface

  42. SAPIEN Engine SAPIEN Engine PhysX Physical Simulator Articulation World Interface Interface ROS Interface Controller Sensor Interface Interface 3D/IMU Force/Joint/Velocity Trajectory Inverse Kinematics Sensor Controller Controller Motion Planning Client API

  43. SAPIEN Engine 3D/IMU Force/Joint/Velocity Trajectory Inverse Kinematics Sensor Controller Controller Motion Planning Client API

  44. SAPIEN Engine Reinforcement Learning 3D/IMU Force/Joint/Velocity Trajectory Inverse Kinematics Sensor Controller Controller Motion Planning Client API

  45. SAPIEN Engine Robot Reinforcement Learning Tasks 3D/IMU Force/Joint/Velocity Trajectory Inverse Kinematics Sensor Controller Controller Motion Planning Client API

  46. SAPIEN Engine

  47. SAPIEN Renderer SAPIEN Engine PhysX Physical Simulator Articulation World Interface Interface ROS Interface Controller Sensor Interface Interface 3D/IMU Force/Joint/Velocity Trajectory Inverse Kinematics Sensor Controller Controller Motion Planning Client API

  48. SAPIEN Renderer SAPIEN Engine SAPIEN Renderer PhysX Physical Simulator Articulation World Interface Interface ROS Interface Controller Sensor Interface Interface 3D/IMU Force/Joint/Velocity Trajectory Inverse Kinematics Sensor Controller Controller Motion Planning Client API

  49. SAPIEN Renderer SAPIEN Engine SAPIEN Renderer PhysX Physical Simulator Renderer Interface Articulation World Interface Interface ROS Interface Controller Sensor Interface Interface 3D/IMU Force/Joint/Velocity Trajectory Inverse Kinematics Sensor Controller Controller Motion Planning Client API

  50. SAPIEN Renderer SAPIEN Engine SAPIEN Renderer PhysX Physical Simulator Renderer Interface Articulation World Interface Interface Rendered ROS Interface Images Controller Sensor Interface Interface 3D/IMU Force/Joint/Velocity Trajectory Inverse Kinematics Sensor Controller Controller Motion Planning Client API

  51. SAPIEN Renderer SAPIEN Renderer Renderer Interface RGBD Normal GLSL Shaders Segmentation

  52. SAPIEN Renderer SAPIEN Renderer Renderer Interface RGBD Normal GLSL Shaders Segmentation Ray Tracing OptiX Shaders

  53. SAPIEN Renderer SAPIEN Renderer Renderer Interface RGBD Normal GLSL Shaders Segmentation Ray Tracing OptiX Shaders 4 spp, 3 bounce, 512x512 OptiX denoise, < 20 FPS

  54. SAPIEN Renderer SAPIEN Renderer Renderer Interface RGBD Normal GLSL Shaders Segmentation Ray Tracing OptiX Shaders Customizable Renderer/Visualizer

  55. SAPIEN Renderer SAPIEN Renderer Renderer Interface Considerations ● Needs to run at real time (~100 FPS) RGBD Normal ○ Rasterizer GLSL Shaders Segmentation ○ Ray tracer denoise? (1spp max) ○ GPU-CPU transfer? (>10 ms) Ray Tracing OptiX Shaders ○ Prebaking? (interaction) Customizable Renderer/Visualizer

  56. SAPIEN Renderer SAPIEN Engine SAPIEN Renderer PhysX Physical Simulator Renderer Interface Articulation World RGBD Interface Interface Normal GLSL Shaders Segmentation ROS Interface Controller Sensor Ray Tracing OptiX Shaders Interface Interface 3D/IMU Force/Joint/Velocity Trajectory Inverse Kinematics Customizable Customizable Sensor Controller Controller Motion Planning Renderer/Visualizer Renderer/Visualizer Client API

  57. SAPIEN Asset SAPIEN Asset ShapeNet PartNet-Mobility Dataset Motion Annotation PartNet 2K models, 14K parts

  58. PartNet-Mobility Annotation Shape cleaning Part segmentation Motion annotation (ShapeNet, 3D Warehouse) (PartNet) (PartNet-Mobility)

  59. Generalize Training Testing

  60. SAPIEN Asset SAPIEN Asset PartNet-Mobility Dataset

  61. SAPIEN Asset SAPIEN Asset PartNet-Mobility Dataset Robot Model

  62. SAPIEN Asset SAPIEN Asset PartNet-Mobility Dataset Robot Model Object Layout

  63. SAPIEN Asset SAPIEN Asset PartNet-Mobility Dataset Robot Model Standard URDF format Object Layout Python API Robot/Scene Builder

  64. SAPIEN SAPIEN Asset SAPIEN Engine SAPIEN Renderer PhysX Physical Simulator Renderer Interface PartNet-Mobility Dataset Articulation World RGBD Interface Interface Robot Model Normal GLSL Shaders Segmentation ROS Interface Object Layout Controller Sensor Ray Tracing OptiX Shaders Interface Interface Robot/Scene 3D/IMU Force/Joint/Velocity Trajectory Inverse Kinematics Customizable Customizable Builder Sensor Controller Controller Motion Planning Renderer/Visualizer Renderer/Visualizer Client API

  65. Task Demonstrations

  66. Task Demonstrations Movable Part Segmentation

  67. Task Demonstrations Movable Part Motion Parameter Segmentation Estimation

  68. Task Demonstrations Movable Part Motion Parameter Segmentation Estimation Part Manipulation

  69. Task Demonstrations Movable Part Motion Parameter Segmentation Estimation Long-horizon Part Manipulation Planning

  70. Task Demonstrations Movable Part Segmentation ● Standard vision problem: detection and segmentation.

  71. Task Demonstrations ● New vision task Motion Parameter Estimation ● Important for control tasks ○ Given the motion parameters, we can use control methods to manipulate the parts

  72. Task Demonstrations ● Manipulation and control Part Manipulation ○ Reinforcement learning ○ Imitation learning ○ robotics control

  73. Task Demonstrations ● Manipulation and control ○ Reinforcement learning ○ Imitation learning ○ robotics control

  74. Task Demonstrations Long-horizon ● Planning Planning ○ Achieve meaningful tasks ○ “Home assistant”

  75. Future Plans/Applications • Benchmark (SAPIEN Challenge) • Vision tasks • Manipulation tasks • Education Platform • Control Algorithms • Robot Learning

  76. Demos

  77. Demos

Recommend


More recommend