How Can Research Teams Collect More Natural Dual-Arm and Hand Data for Robot Learning?

Embodied AI systems learn from interaction. To teach a robot how to pick, place, assemble, sort, or use tools, researchers need demonstration data that describes not only the final result, but also the sequence of movements that produced it.

This becomes especially challenging in dual-arm and dexterous manipulation. The operator must coordinate two arms, adjust joint positions continuously, and often control individual fingers at the same time. Conventional controllers can operate a robot, but they may require the user to translate a natural human action into buttons, joysticks, or separate commands. Video can record what happened, yet it does not directly provide complete joint-level motion data.

The ExoArm-7 Pro wearable exoskeleton is designed to address this gap by capturing human arm movements directly and turning them into high-frequency data for robot teleoperation, imitation learning, embodied AI, and VLA development.

Why Is Bimanual Demonstration Data Difficult to Collect?

Single-arm tasks already involve position, orientation, timing, and contact. With two arms, the data-collection problem becomes more complex because both sides must remain coordinated throughout the task.

Consider a few common examples:

• One hand holds a container while the other opens it
• Two arms lift or reposition a large object together
• One hand stabilizes a component while the other uses a tool
• An object is transferred from one hand to the other
• Both hands organize or fold a flexible item

For these demonstrations to be useful, the system must capture subtle joint changes with low delay. It must also remain comfortable enough for operators to repeat a task many times without the wearable device significantly changing their natural movement.

How Does ExoArm-7 Pro Capture Human Arm Motion?

ExoArm-7 Pro uses a 14DOF dual-arm configuration consisting of seven passive joints on each arm. When the operator moves, the wearable structure follows the motion and records the corresponding joint positions.

Each joint uses an in-house-developed encoder module, while 14-bit joint data provides fine-grained position feedback. This allows the system to detect small angular changes instead of recording only broad gestures.

For a teleoperation workflow, the captured positions can be mapped to a robotic arm so the robot follows the operator’s movement. For a learning workflow, the same motion information can become part of a structured human-demonstration dataset.

Can It Respond Quickly Enough for Real-Time Teleoperation?

Responsiveness is essential when an operator needs to correct a robot’s motion while interacting with an object. If the data arrives too slowly, the robot may feel disconnected from the operator and fine manipulation becomes harder to control.

ExoArm-7 Pro supports sampling rates of up to 1000 Hz together with millisecond-level low-latency control. This high-frequency data stream helps preserve rapid changes in arm position and supports more responsive human-to-robot motion mapping.

Dual-IMU sensing adds motion and orientation information beyond the joint encoders. The standard system includes both a back-mounted IMU and an external IMU, providing additional data for interpreting the operator’s movement and orientation.

Will a Wearable Exoskeleton Restrict the Operator?

A motion-capture system should follow the user without becoming the focus of the task. If it is too heavy or uncomfortable, the operator may change posture, shorten movements, or become less consistent during repeated demonstrations.

ExoArm-7 Pro weighs 2.6 kg and uses a lightweight ergonomic structure. Each encoder joint module weighs less than 70 g while supporting up to 10 N·m of load-bearing torque. Adjustable mechanical limits allow the joint modules to be configured for the wearable structure and operating requirements.

This design makes the system more practical for repeated data-collection sessions in laboratories, research centers, and robotics development environments.

What If the Task Requires Finger-Level Manipulation?

Arm trajectories can guide a robot toward an object, but they do not describe how the fingers shape a grasp or manipulate a tool. Tasks involving dexterous hands therefore require another layer of motion data.

ExoArm-7 Pro can be paired with the optional ExoGlove Pro teleoperation data gloves. The gloves use five VetraSense sensors to reconstruct a complete 25DOF hand pose, reducing the number of sensors required while preserving detailed finger-motion information.

Each fingertip supports 6DOF tracking, enabling the system to record both position and orientation. Adjustable vibration feedback can also be used to provide the operator with interaction cues during teleoperation.

ExoGlove Pro outputs data at up to 180 Hz and supports wired latency below 10 ms. Its SDK and motion-retargeting algorithms help translate human hand poses to multiple dexterous-hand and robotic platforms.

The gloves are particularly relevant for datasets involving:

• Precision grasping
• Tool operation
• Component assembly
• Object rotation and reorientation
• Flexible-object manipulation
• Coordinated arm, hand, and finger motion

ExoGlove Pro is sold separately and is not included in the standard ExoArm-7 Pro package. Teams that only require arm-level data can use the standard system, while projects involving dexterous manipulation can add the gloves as an expansion.

How Can the System Fit an Existing Robot Platform?

Robot-learning teams rarely use identical hardware and software stacks. One laboratory may work with a research arm, another with a mobile dual-arm robot, and another with a custom dexterous hand.

ExoArm-7 Pro includes an official SDK and free integration support to help developers connect the captured data to their own platforms. The visual software interface supports Windows, macOS, and Linux.

Custom control handles and an optional glove adapter provide additional flexibility for different operating methods. For hand-based applications, multi-platform SDK support and motion-retargeting algorithms help adapt the human motion data to robots with different joint structures.

This means teams can treat ExoArm-7 Pro as part of a broader teleoperation and data pipeline rather than as a closed system tied to one robot.

Where Can the Collected Data Be Used?

The same wearable workflow can support several stages of robotics development.

Real-Time Robot Teleoperation

An operator can control dual robotic arms through natural upper-limb motion, reducing reliance on separate joystick commands for every joint.

Human Demonstration Collection

Researchers can record how a person performs a task, including the timing and coordination between both arms.

Imitation Learning

Repeated demonstrations can be organized into datasets that help a robot learn manipulation policies from human examples.

VLA Dataset Development

Arm and optional hand-motion data can be combined with other sensor inputs and task instructions as part of a Vision-Language-Action research pipeline.

Robotics Education

Universities and laboratories can use the system to study motion retargeting, human-robot interaction, teleoperation, and embodied intelligence within one development platform.

Choosing the Right Configuration

The standard ExoArm-7 Pro package includes the wearable exoskeleton, a back-mounted IMU sensor, and an external IMU sensor. This configuration is suitable for teams focused on dual-arm motion capture, robot-arm teleoperation, and upper-limb demonstration data.

Projects that require detailed grasping and finger trajectories can add the optional ExoGlove Pro package, which includes two motion-capture gloves, a wireless receiver and battery module, and two USB-C data cables.

Choosing between the two configurations depends on the required action resolution. Arm-level tasks may not need glove data, while dexterous manipulation and complete arm-to-hand demonstrations benefit from capturing both layers together.

Turning Natural Human Action into Structured Training Data

The central challenge in embodied AI is not simply collecting more data. It is collecting data that accurately represents how people perform useful physical tasks.

By combining 14DOF dual-arm tracking, high-resolution encoder feedback, sampling rates up to 1000 Hz, dual-IMU sensing, and optional 25DOF hand-pose capture, ExoArm-7 Pro provides a practical path from natural human movement to structured teleoperation and robot-learning data.

For teams building dual-arm robots, dexterous manipulators, imitation-learning systems, or VLA models, this wearable approach can make demonstrations more intuitive for operators and more informative for the machines learning from them.