Our lab

Where human dexterity becomes data.

Where human dexterity becomes data.

Where human dexterity becomes data.

The Kavana lab is where we capture how skilled hands really work: every joint angle, every contact, every newton of force, together with the neuromuscular signals that drive them.

People perform real manipulation tasks while we record them with the combination of sensors each task calls for: motion capture, Mudra™ EMG, egocentric video, force sensing and tactile gloves. We turn those sessions into multimodal datasets for training dexterous robotic hands.


The Kavana lab is where we capture how skilled hands really work: every joint angle, every contact, every newton of force, together with the neuromuscular signals that drive them.

People perform real manipulation tasks while we record them with the combination of sensors each task calls for: motion capture, Mudra™ EMG, egocentric video, force sensing and tactile gloves. We turn those sessions into multimodal datasets for training dexterous robotic hands.


A researcher in the Kavana capture volume wearing a head-mounted POV camera and a Mudra band, surrounded by OptiTrack cameras

What we build

Multimodal datasets for dexterous manipulation.

Multimodal datasets for dexterous manipulation.

Multimodal datasets for dexterous manipulation.

The lab exists to produce datasets that combine paradigms no single sensor can cover. Each one targets a core problem in five-finger manipulation.

01

Joint-angle estimation

Ground-truth finger kinematics from optical motion capture, paired with EMG so hand pose can be learned from the wrist.

02

Computer-vision uplift

Labels that teach vision models what cameras miss: hands through occlusion, glare, low light and fast motion.

03

Applied-force estimation

Grip and contact force measured in newtons, aligned with the muscle activity that produced it.

The capture rig

Five modalities. The right mix for every task.

Five modalities. The right mix for every task.

Five modalities. The right mix for every task.

Each modality sees a different side of the hand, from the outside view of cameras to the inside view of muscle activity. For every study we combine the ones the task calls for, so each signal can supervise the others.

Motion capture

12-camera OptiTrack volume

Marker-based tracking of hand, fingers and objects, giving precise kinematic ground truth for pose.

Neuromuscular

Mudra Ultimate

8 EMG channels at 500–4000 Hz, 24-bit, with IMU and PPG, streaming over Wi-Fi. Intent and effort, read at the wrist.

Point of view

Egocentric RGB / RGB-D

Head-mounted RGB or depth cameras such as ZED record the task from the operator’s eyes, the way a robot will see it.

Force

Force sensors

Instrumented objects and surfaces measure applied force, so every grasp, press and twist comes with a number.

Tactile

Tactile glove array

A pressure-sensing array across the palm and fingers maps where, and how hard, the hand touches.

A hand fitted with motion-capture markers and a Mudra band, with an OptiTrack camera behind it

The foundation

One clock for every signal.

One clock for every signal.

One clock for every signal.

A multimodal dataset is only as good as its alignment. The lab is built on a synchronization layer that uses proven, open methods such as Lab Streaming Layer (LSL).

Whichever streams a task uses, from the camera array to the EMG channels to the force and tactile sensors, are timestamped on a shared clock. A muscle burst, a finger flexion and a contact force line up in the same moment.

A Kavana researcher monitoring synchronized EMG and motion-capture streams

12

12

OptiTrack cameras

8

8

EMG channels

5

5

Modalities to combine

1

1

Shared clock

Research in parallel

Many studies. One rig.

Many studies. One rig.

Many studies. One rig.

Several research programmes run in the lab at the same time, sharing the same capture infrastructure, calibration and data formats. What one study records, every other study can build on.

A participant recorded in the capture volume while an operator runs the session
A participant with hand markers and a Mudra band reaching toward a live display
A researcher fitting a POV camera headset on a participant inside the OptiTrack cage

Why it matters

We believe synchronized multimodal data, combined with 12 years of neural-signal know-how, can become the industry standard for training robotic hands to do what wasn’t possible before.

We believe synchronized multimodal data, combined with 12 years of neural-signal know-how, can become the industry standard for training robotic hands to do what wasn’t possible before.

We believe synchronized multimodal data, combined with 12 years of neural-signal know-how, can become the industry standard for training robotic hands to do what wasn’t possible before.

Fine-grained, dexterous object interaction, learned from the people who do it best.

Work with the lab