QuickMagic Robot Motion Capture from Video: Unitree G1, H1 and H1-2 File Guide

1Record or generate a useful source motion

• Start with a short, slow action containing the key contact transition.
• Keep the full body and feet visible for locomotion or balance tasks.
• Use adequate light and low motion blur.
• Avoid severe self-occlusion and multiple-person overlap.
• Show the floor when foot placement matters.
• Use text-to-motion for behavior ideas; use video when exact human timing matters.
• Avoid beginning with kicks, falls, jumps or fast spins as the first robot test.

2Select the exact UniRobot preset

Choose the model shown in the current QuickMagic export interface and record the selection with the project. If the target is a modified G1 EDU, custom hand or altered waist configuration, verify whether the preset matches the actual URDF and controller state vector.

3Validate the exported motion file

Minimum validation questions

• Which robot and exact variant does this file target?
• How many controlled joints are present?
• What is the joint-name and column order?
• Are joint values radians or degrees?
• Are linear positions meters or centimeters?
• Which axis is up, forward and left?
• Is the base orientation Euler, quaternion or another representation?
• If quaternion: is the order wxyz or xyzw?
• What are the frame rate and time step?
• Are contact labels, joint velocities or confidence values included?

4Retarget and constrain the reference

Kinematic retargeting

• Map human body segments to robot links and joint axes.
• Fit or scale the human reference to robot proportions.
• Clamp joint positions, velocities and accelerations.
• Define a neutral/base pose compatible with the controller.
• Preserve important end-effector trajectories without forcing impossible poses.

Contact and base handling

• Detect or define planted-foot intervals.
• Project foot targets to the intended terrain.
• Adjust pelvis/base trajectory to maintain feasible support.
• Check self-collision and environmental collision.
• Treat hand-object interaction as a separate constraint problem.

5Validate in simulation and train the tracker

Kinematic validation

• Every frame loads without missing or duplicated joints.
• The neutral pose is correct.
• Left/right legs and arms are not swapped.
• Joint limits and self-collision are respected.
• Duration matches the original reference.

Dynamic validation

• The robot remains balanced under gravity.
• Foot contacts do not penetrate or slide excessively.
• Requested torque, velocity and acceleration stay within limits.
• The controller recovers from small disturbances and model error.
• Performance remains stable with realistic latency, friction and sensor noise.

Unitree RL MjLab separates motion-reference preparation, training, simulation playback and physical deployment. This separation is the correct mental model for QuickMagic robotics data as well.

6Sim2Real deployment and physical safety

Before a physical test, confirm the official development-computing configuration, firmware and SDK version, network interface, robot mode and controller assumptions. Use the vendor's current documentation rather than copying settings from an unrelated robot or software version.

• The exact physical robot matches the simulated URDF/USD and policy configuration.
• Emergency stop and remote stop have been tested.
• The test area is clear and access-controlled.
• The robot begins in the vendor-prescribed safe state.
• The first motion is slow, near-upright and short.
• No people are inside the fall, swing or kick area.
• Logs capture state, reference, action, latency and safety events.
• A qualified operator can stop the experiment immediately.