Primate | Teach robots with your hands

Product

The worker wears it. The robot runs it.

While you wear the gripper on shift, it records what the cameras see, where your hand moves, and when the fingers open and close. Then the same gripper attaches to the arm and the robot performs the job.

Human hand holding the gripper on the left, identical gripper mounted on a robot arm on the right

Handheld gripper

Wear it to teach the robot.

Press one button on the gripper to record:

Its position and orientation
What the two cameras see
When the fingers open and close

Robot gripper

Attach it to the robot to do the job.

Same device the worker wore
Performs the task it learned
Works with any robot

How it works

From human skill to robot skill.

The gripper records the workers' normal tasks, and the robot learns from that data.

Record

The worker wears the gripper and does their normal job. One click on the platform starts recording, and every frame captures what the cameras see, the position and orientation of the gripper, and how open the fingers are. See what is recorded ↓

Review

The worker watches the recordings on the platform and scores them like a performance review. They mark errors, name each step, and rate how well the task was done. No AI expertise is needed. They only need the judgment they use at work. See the review screen ↓

Train

The worker clicks on the platform and trains the robot on our cloud infrastructure, so the customer needs no expensive hardware for it. Roughly 500 recordings are enough for one task. See the proof ↓

Deploy

The worker clicks start on the platform and the robot does the task it learned, on an off-the-shelf arm (ABB, KUKA, UR, etc.) at the customer site. We bring the PC for the AI, and the arm too if the customer does not own one.

Improve

The robot does the task, and when it occasionally makes a mistake the worker marks what went wrong on the platform and clicks to train it again. The robot becomes more reliable and faster with every round.

Recording in action

Left gripper cameraiWhat the left gripper's camera sees.

Right gripper cameraiWhat the right gripper's camera sees.

Worker viewiThe worker doing the task, from the outside.

3D positioniPosition and orientation of each gripper in 3D, live.

Left gripper camerai

What the left gripper's camera sees.

Right gripper camerai

What the right gripper's camera sees.

Worker viewi

The worker doing the task, from the outside.

3D positioni

Position and orientation of each gripper in 3D, live.

Review screen

Quality ratingiRate each recording, like grading a person's work.

LabelsiLabel each action so the robot learns the sequence.

Step bariMark the start and end of each step so the robot learns the sequence.

Error bariMark where the task went wrong.

Quality ratingi

Rate each recording, like grading a person's work.

Labelsi

Label each action so the robot learns the sequence.

Step bari

Mark the start and end of each step so the robot learns the sequence.

Error bari

Mark where the task went wrong.

One place to record, review, train, run, and improve.

Why Primate

Four ways to teach a robot. Three fail in real work.

Some methods train the robot in a clean copy of the world, not on the factory floor. Others train it on a human hand, not its own gripper. Either way, it learns the job in a place and a body that are not its own.

Method	How it works	Reality check
Simulation	The robot is trained in a physics simulator that approximates objects, materials, contacts, and dynamics.	Deformable materials like cloth, foam, food, liquids, and phenomena such as fracture, are expensive or impossible to simulate accurately. For every element in the simulation, the user must set exact parameters such as friction, stiffness, and mass. Those parameters most likely differ from reality.
Teleoperation	A person controls the robot arm with a joystick and performs a task with it.	Because the operator uses a joystick instead of their own arm, the motion is unnatural and the operator lacks full control, so they move slowly and carefully to prevent damage, and the robot inherits this caution. Because the operator must control the robot directly, one robot needs one operator for as long as data is collected.
Learning from video	The robot is trained on video of human hands from public footage or a body-mounted camera.	The human hand is far more complex than any robot hand, so there is an embodiment gap even with five-fingered robot hands. Objects also occlude the fingers during manipulation, so inferring the fingers' poses is challenging. Moreover, videos will never provide force or tactile data, and this is a modality problem. See it happen ↓
Primate Gripper	A worker wears the robot's own gripper and sensors while doing the actual job.	There is no embodiment gap, because the training data comes from the same hand and sensors the robot will use. The work is the customer's real task during real production, not a mockup, not a proxy, and not a simulation.

Roadmap

Beyond the gripper.

The gripper already handles tasks like pick-and-place and assembly. The glove is the next step: it gives the robot dexterity closer to a human hand and unlocks new tasks without adding delays. The worker wears one glove and the robot wears its twin. Both use the same kinematics, sensors, and cameras, so skills transfer directly from the human hand to the robot hand.

Once the hands are dexterous enough, the robot also needs to move around the plant. A few unobtrusive sensors attach to the worker's normal clothes, forming a sensor suit that needs no special gear. The suit records whole-body movement during normal tasks. It captures how the worker coordinates moving and manipulating objects, without slowing the work or changing the routine. That data is what humanoid robots need to walk, balance, handle objects, and operate in human spaces.

A worker in a sensor suit walks while a humanoid robot mirrors the same motion

Prototype sensor glove with blue backing plates and wrist module

Team

Built by roboticists.

Luigi Campanaro

Co-founder & CEO

PhD, Robotics & AI · University of Oxford.
Postdoctoral Researcher · EPFL.
Designed one of the first three reinforcement-learning locomotion controllers for legged robots in the UK.
Built the AI that automated phone loading into chargers for a high-volume production line, completing 550 consecutive cycles without error. Watch the video ↓

Locomotion controller · RL demo ↗

Axel Praplan

Co-founder & CTO

MSc in Robotics · EPFL.
Robotics Research Engineer · EPFL.
Built a full tendon-driven humanoid robot from scratch: mechanical design, actuation, control, integration.
Deep expertise: hardware design, mechatronics, rapid prototyping.

Tendon-driven humanoid · built from scratch ↗

Advisors

Prof. Daniele De Martini

Professor of Robotics

University of Oxford

Co-leads the Mobile Robotics Group at Oxford alongside Prof. Paul Newman (Founder and CEO of Oxa). His research on robotic perception and autonomy enables robots to navigate and interpret the world in the most challenging conditions.

Prof. Perla Maiolino

Professor of Robotics

University of Oxford

Deputy Director of the Oxford Robotics Institute. She developed CySkin, one of the few existing large-area robotic skins, giving robots the sense of touch that makes safe, intelligent manipulation possible.

Alf Müller

Entrepreneur in Digital Commerce

Co-owner, Purecrea GmbH / Bastelgarage.ch

Swiss entrepreneur and maker, co-owner of Purecrea GmbH behind Bastelgarage.ch, one of Switzerland's largest online stores for DIY electronics, robotics and maker products. He combines e-commerce with hands-on engineering, rapid prototyping and robotics, electronics and AI.

Wearable data collection for task automation

The worker wears it. The robot runs it.

From human skill to robot skill.

Collect data. Or drive the robot directly.

Four ways to teach a robot. Three fail in real work.

Serious work, simple gripper.

Beyond the gripper.

Built by roboticists.

Turn your work into robot skills.

Wearable data collection for task automation

The worker wears it. The robot runs it.

From human skill to robot skill.

550 consecutive runs, zero errors.

The robot can only do what the camera sees.

Collect data. Or drive the robot directly.

Four ways to teach a robot. Three fail in real work.

Serious work, simple gripper.

Beyond the gripper.

Built by roboticists.

Turn your work into robot skills.