Imitation Learning for Humanoid Robots

Refer to Imitation Learning for more details.

Four possible source of demonstrations for humanoid robots:

• policy execution on hardware (robot)

  • collecting data on physical robots (requires a laborious setup of the environment and raises significant safety concerns)

  • collecting data in simulation (sim-to-real gap)

• Teleoperation (robot)

  Control flow for learning from teleoperated demonstrations:

  1. human operator (VR, Exoskeleton, Optical Tracking, Motion Capture, Joystick, ALOHA)

  2. motion retargeting

  3. desired robot trajectory

  Some limitations:

  • a majority of teleoperation systems capture only manipulation skills, full-body sensing, including human gaits are missing (requires IMU or exoskeletons which are expensive)

  • the teleoperation data may limited if the robot’s kinematics do not enable seamless retargeting

  • time-consuming to scale

• motion capture from human (3D) (human)

  Captures humans interacting with various objects while moving around.

  • require heavily instrumented environments and actors

  • less outdoor activities

• human videos from internet (2D) (human)

  Obtain rich and diverse human motion datafrom the internet.

  • lower quality, containing noise, non-physical artifacts