Teleoperation data collection
Human operators control robot arms while joint positions, velocities, and force-torque readings are recorded — producing action-labeled trajectories for behavior cloning and diffusion policy training.
Definition
Teleoperation data collection is the process of having a human operator remotely control a robot while recording every aspect of the interaction. The operator provides the intelligence — deciding what to grasp, how to approach, where to place — while the robot's own sensors record the resulting joint positions, velocities, torques, and end-effector poses.
The key advantage: teleoperation data is already in the robot's action space. There is no retargeting step, no coordinate transform, no embodiment gap. The recorded trajectory is exactly what the robot needs to replay or generalize from.
This makes teleoperation data the most direct input for behavior cloning, diffusion policy training, and action-chunking architectures. Every demonstration is a ground-truth trajectory that can be fed straight into the model without any additional processing or domain adaptation.
Why it matters
Data is recorded in the robot's own joint space. No kinematic retargeting, no embodiment transfer. The training signal matches the deployment configuration exactly — what the robot records during teleoperation is what it will execute during autonomous operation.
Every timestep comes with a ground-truth action label — the control signal the operator sent. No post-hoc annotation needed for behavior cloning. The action supervision is built into the collection process itself, reducing pipeline complexity and annotation cost.
One teleoperation setup can collect data for dozens of different tasks by changing the task protocol. The hardware stays the same; only the instructions change. This makes it practical to build multi-task datasets without reconfiguring the capture infrastructure.
What gets recorded
Full joint state at 100+ Hz. Every degree of freedom recorded with sub-millisecond timestamps for precise trajectory reconstruction.
Position and orientation of the gripper or tool in Cartesian space. Computed via forward kinematics and verified against external tracking.
Open/close position, grip force, and contact events. Critical for learning grasp timing and force-sensitive manipulation tasks.
6-axis force-torque sensor readings at the wrist. Captures contact forces, insertion resistance, and load changes during manipulation.
Wrist-mounted and external camera feeds with synchronized color and depth. Multiple viewpoints for multi-view policy architectures.
Raw operator commands from the teleoperation interface — joystick axes, VR controller poses, or leader arm joint states.
Interfaces
Kinesthetic teaching with matched leader-follower pairs. Natural force feedback, intuitive for operators, high-bandwidth control. The leader arm mirrors the follower's kinematics, so the operator feels the task dynamics directly — producing smoother, more natural demonstrations.
Spatial tracking via VR headsets and hand controllers. Allows remote operation without physical proximity to the robot. Operators see the robot's camera feeds in the headset and control end-effector pose through natural hand movements in 6-DoF space.
Joystick, SpaceMouse, haptic devices, and custom input hardware. We adapt to whatever control interface your robot platform uses. If your team has an existing teleoperation stack, we integrate with it rather than replacing it.
Data Explorer
Teleoperation datasets collected by Humaid are browsable through the robotics data explorer. Review individual teleoperation sequences with synchronized video playback, verify action segmentation labels against recorded trajectories, and inspect the full metadata for each recording — including joint positions, velocities, and force-torque profiles.
Each sequence is downloadable in multiple formats. Teams can access raw MCAP sensor recordings, processed annotations in JSON, and rendered video overlays — all without writing custom data parsing scripts. Open the data explorer (opens in a new tab).
Get started
Tell us your robot platform, control interface, and target tasks. We will configure the teleoperation setup, train operators, and deliver synchronized trajectory datasets ready for your training pipeline.