MANUS Gloves for Robotics: Precision Hand Data for AI Training & Teleoperation

Most discussions about robot learning focus on model architectures, compute, and simulation environments. The conversation rarely lands on what actually limits progress in dexterous manipulation: the quality of hand data used to train and control robot hands.

Vision-based hand tracking — cameras, depth sensors, computer vision — breaks down precisely where robotics needs it most. When fingers overlap, when the hand rotates away from the camera, when a grasp occludes the fingertips, tracking fails. For casual interaction this is acceptable. For robot policy training, where a single corrupted joint angle can degrade an entire demonstration, it is not.

MANUS data gloves solve this problem at the source. By tracking hand articulation electromagnetically — without cameras, without line-of-sight requirements, without drift — they deliver the clean, complete hand motion data that robotics pipelines actually need. This article explains how they work, where they fit in a robotics stack, and how to get started. Available through Knoxlabs as individual units or as part of a fully configured robotics stack.

Why Hand Data Is the Robotics Bottleneck

Humanoid robot development has accelerated significantly in recent years. Full-body motion capture for locomotion, gait, and posture is well-solved — Xsens suits and similar IMU-based systems have been deployed in robotics for over a decade. Arms and torso kinematics are tractable problems.

The hand is different. Human hands have 27 bones, over 30 joints, and execute manipulation behaviors — pinching, grasping, rotating, palpating — that require millimeter-level precision to capture meaningfully. The failure modes of hand tracking matter enormously:

• Adjacent finger occlusion — when index and middle fingers overlap during a grasp, optical systems cannot distinguish individual joint angles
• Wrist rotation drift — IMU-based systems accumulate error during extended operation; drift corrupts long demonstration sessions
• Depth ambiguity — RGB-D cameras cannot resolve the depth of curled fingers with sufficient precision for fine manipulation
• Fingertip resolution — the distal joints of each finger, where most contact interaction happens, are the hardest to track optically

The consequence is a training data quality problem. Imitation learning and reinforcement learning from human demonstrations are only as good as the demonstrations themselves. Noisy or incomplete hand data produces robot policies that fail to generalize.

This is the gap MANUS was built to fill — and why leading research teams consistently arrive at hardware-based glove tracking when they need production-quality hand data.

How MANUS Tracks the Hand

MANUS Metagloves use electromagnetic field (EMF) tracking rather than optical or inertial methods. Small sensors embedded in the glove measure their position and orientation relative to an electromagnetic field generated by a body-worn transmitter. This approach eliminates the two core failure modes of competing technologies:

• No cameras — there is nothing to occlude. Fingers can overlap, curl, or face any direction without affecting tracking quality
• No IMU drift — EMF measurements are absolute, not accumulated. Data quality is consistent at the start and end of an 8-hour session

The result is 25 degrees of freedom per hand — every metacarpal, proximal, middle, and distal joint — tracked continuously at millimeter precision.

MANUS Core — the companion software — processes raw sensor data and outputs three simultaneous streams:

• Raw Skeleton — anatomical hand model, all 25 DoF, in real time
• Retargeted Skeleton — human hand kinematics mapped to your specific robot hand's joint configuration
• Sensor Data — raw electromagnetic measurements for custom processing pipelines

These streams are available via the MANUS SDK (C++, Windows and Linux), ROS 2 topics, and direct integration with NVIDIA Isaac Lab. You choose the integration path that fits your stack.

Use Case: AI Training Data Collection

Robot policy training via imitation learning requires human operators to demonstrate tasks while the system records the motion. The quality of those demonstrations directly determines the quality of the trained policy. MANUS gloves are increasingly the standard input device for this workflow because they meet the data quality bar that robot learning pipelines require.

What robot learning pipelines need from hand data

Policy training systems — whether via behavioral cloning, diffusion policy, or reinforcement learning from demonstrations — are sensitive to:

• Joint angle completeness (missing joints create systematic errors in learned behaviors)
• Temporal consistency (dropout frames corrupt demonstration quality disproportionately)
• Fingertip precision (the distal joints execute the actual manipulation; errors here compound into policy failures)

MANUS addresses all three. Occlusion-free EMF tracking means no missing joints during complex grasps. Drift-free measurement means consistent quality across long recording sessions. And 25 DoF coverage includes the distal joints that matter most for contact-rich tasks.

NVIDIA Isaac Lab 2.3 native integration

MANUS gloves are natively supported in NVIDIA Isaac Lab 2.3 as a teleoperation and data collection device. Operators put on the gloves, open Isaac Lab, and the human hand configuration maps directly to simulated robot joint positions in real time. Demonstrations recorded in Isaac Lab feed directly into Isaac Lab Mimic for augmentation and scaling before policy training.

This makes MANUS the practical choice for research teams working within the NVIDIA robotics ecosystem — there is no custom integration required. The complete Isaac Lab setup guide is in Article 3: MANUS + NVIDIA Isaac Lab Integration.

For teams collecting data outside of Isaac Lab, MANUS exports to CSV and FBX — standard formats compatible with PyTorch, TensorFlow, and most ML data pipelines.

Use Case: Real-Time Teleoperation

Teleoperation demands a different performance profile than data collection. Where training data can be post-processed, teleoperation is unforgiving: latency above approximately 50ms is perceptible to operators, degrading control quality. Any tracking dropout causes control loss. And for hazardous environment applications — nuclear facilities, chemical labs, remote surgical tools — control fidelity is a safety requirement.

MANUS is designed to meet these demands. The C++ SDK streams hand data via ROS 2 with latency under 7ms. The retargeting pipeline in MANUS Core handles the mapping from human hand geometry to robot hand joint configurations in real time, accounting for scale differences and mechanical constraints.

Haptic feedback closes the control loop

The Metagloves Pro Haptic adds one vibrotactile actuator per finger, with 256 modulation channels per motor. When the robot hand makes contact with an object, the haptic layer signals that contact back to the human operator's corresponding finger.

This matters practically. Without haptic feedback, operators over-squeeze objects or under-squeeze without knowing it. With it, experienced operators adapt within minutes and achieve manipulation performance that approaches direct hand operation.

At NVIDIA GTC 2026, Dr. Scott Walter — a pioneer in humanoid robotics — used MANUS gloves to teleoperate the Sharpa Wave, a 22-DoF dexterous robotic hand, in real time. The demonstration included catching a thrown ball — a task requiring sub-100ms latency and precise multi-finger coordination. He called MANUS his "favorite data glove." The complete teleoperation stack breakdown — layers, retargeting, ROS 2 integration — is in Article 5: Teleoperation Stack 101.

Choosing Your Glove: Pro vs Pro Haptic

MANUS makes two primary gloves for robotics applications. The right choice depends on your use case:

For teams focused entirely on data collection and AI training pipelines, the Metagloves Pro is the right choice. The haptic layer adds cost and complexity that does not contribute to data quality. For teams doing live teleoperation — especially in applications where contact sensing matters (surgical, hazardous, precision assembly) — the Pro Haptic is worth the addition.

Best for AI training

Metagloves Pro

25 DoF · EMF · <7ms latency · CSV/FBX export · ROS 2 · NVIDIA Isaac Lab 2.3

View on Knoxlabs

Best for teleoperation

Metagloves Pro Haptic

25 DoF · Per-finger haptics · 256-channel actuation · ROS 2 · NVIDIA Isaac Lab 2.3

View on Knoxlabs

Both are available through Knoxlabs. See the full robotics component catalog for configuration options, or see Article 2 for how both gloves integrate with Xsens full-body motion capture.

Supported Integrations

MANUS is designed for integration, not isolation. The supported software and middleware ecosystem covers the major robotics development environments:

The MANUS C++ SDK exposes the full data stream on both Windows and Linux. For ROS 2 deployments, MANUS publishes hand skeleton data as standard ROS topics — no custom message types required for basic integration.

When paired with an Xsens Link full-body suit, the combined system outputs synchronized whole-body + hand kinematics — the complete human motion input layer for humanoid robots. That integration is covered in Article 2: MANUS + Xsens.

Getting Started with Knoxlabs

MANUS gloves are available through Knoxlabs as individual units or as part of a configured robotics stack. As an authorized MANUS reseller, Knoxlabs sources, configures, and ships alongside any other component of your teleoperation or data collection system — Xsens suits, VR headsets, robot hands — as a single order.

For teams that need hardware delivered ready for their pipeline rather than requiring provisioning work, Knoxlabs handles configuration and pre-shipment verification at the Glendale, CA facility. For the full picture of what that process looks like end-to-end, see From Vision to Reality: White-Glove XR Deployment That Actually Works and the robotics-specific deployment guide in Article 6.

The Knoxlabs Robotics & Teleoperation hub has the full component catalog with all hardware options across every stack layer.