The operator moves a master device (mocap glove, exoskeleton, controller, leader arm). Those motions are sensed, mapped into the robot's task or joint space and transmitted over a communication link (often low-latency, e.g. via WebRTC) to a follower robot that reproduces them. The robot's sensors and cameras stream video back and — in bilateral setups — forces and torques, closing the feedback loop. The state-action trajectories recorded during teleoperation form demonstration datasets on which autonomous manipulation policies are trained via behavior cloning / imitation learning.
It lets humans perform dexterous physical tasks where robot autonomy is insufficient or human presence is dangerous, while also solving the data-scarcity problem: it supplies human-driven demonstrations for training manipulation policies that cannot easily be scraped from the internet.
Input device capturing the human operator's motions and intent: mocap glove, exoskeleton, controller, leader arm or VR headset.
Channel transmitting control signals and telemetry between operator and robot with the lowest possible latency (e.g. WebRTC).
Robot reproducing the operator's motions in the remote environment, equipped with actuators, a gripper/end-effector and sensors.
Return of video, forces and torques to the operator, providing a sense of contact with the remote environment.
Official
Communication-link delays degrade bilateral control stability and telepresence quality.
Kinematic and scale differences between human and robot make faithful motion retargeting hard.
Raymond Goertz at Argonne National Laboratory builds an electrical master-slave manipulator with bilateral force feedback — the birth of modern teleoperation.
Commercialization of teleoperated robotic surgery brings teleoperation beyond the laboratory.
Low-cost bimanual teleoperation rigs (ALOHA) enable collecting demonstrations to learn precise manipulation policies from minutes of data.
Vision-Language-Action models trained partly on teleoperation-collected robot trajectories fuse web knowledge with robot control.
Round-trip delay in the operator-robot loop; critical for stability and telepresence.
How operator motion is mapped onto the robot's task/joint space.
Type of feedback provided: visual, haptic (force/torque), bilateral.
The robot's actions are entirely conditioned on the operator's real-time input.
Control is inherently sequential and real-time; parallelism appears only in later training on the collected data.
