Training humanoid robots to manipulate objects requires large volumes of demonstration data, and that data must be precise. Every dropped grasp or misaligned finger in the training set becomes a failure mode in deployment.
N1 Robotics built Waldo to close that gap. Waldo is a humanoid teleoperation platform that translates natural operator hand motion into structured manipulation trajectories, ready to feed directly into training pipelines. Doing that reliably at scale meant the hand tracking input had to capture operator intent exactly, not approximate it.
MANUS gloves are professional data gloves built on EMF tracking. They deliver millimeter-level hand tracking across 25 degrees of freedom with no occlusion and no drift. Every joint, every micro-movement, every pinch geometry is recorded as it happens.
For N1 Robotics, this level of fidelity is the starting point for everything downstream. Waldo uses MANUS gloves as the operator interface, and MANUS Core prepares that hand motion data into a form ready for retargeting onto the robot hand. Because the grasp intent and finger coordination are preserved through capture and structuring, Waldo has accurate input to work with, and the resulting demonstration reflects what the operator actually did.
Capturing hand motion with precision is necessary but not sufficient. The harder challenge is retargeting: mapping what a human hand does onto a robotic hand that has different proportions, different joints, and different mechanical constraints.
N1 Robotics addressed this by building WaldoRT, a custom retargeter that maps MANUS output directly onto a robot hand model. WaldoRT preserves grasp topology, finger coordination, and pinch geometry across different operator hand sizes and different robot end effectors. It requires no manual per-operator or per-end-effector retuning. As long as the URDF is available, WaldoRT can generate the retargeting configuration for that hand. The system has been tested across multiple platforms including the Inspire, BrainCo Revo 2, Sharpa, and Dexhand, among others.*
The result is a retargeting layer that scales with the hardware. The same natural human grasp maps cleanly to any supported robotic hand, which means the fidelity built into the MANUS signal carries through to the final demonstration. That consistency is what makes the numbers in practice meaningful.
N1 Robotics tested Waldo against a Meta Quest-based teleoperation setup using the same operator, the same three manipulation tasks (shape insertion, grabbing toast, and cup stacking with a BrainCo Revo2), and the same ten-minute window.
Across all three tasks, Waldo with MANUS gloves produced more attempts, more successful demonstrations, and faster episode times. The gap was most pronounced in shape insertion, where Meta Quest operators averaged300 seconds per episode at an 18.2 percent success rate against Waldo's 42.86seconds at 93.3 percent — a difference that held consistently across grabbing toast and cup stacking as well.
Part of this difference comes from the nature of the interface itself. Waldo operators can instantly re-attempt a failed grasp without external assistance. Meta Quest operators required a second person to reset the environment. When the goal is collecting demonstrations at scale, the ability to iterate quickly within a session matters as much as the quality of each individual attempt.
* This is available for the Metagloves Pro users.
