Robotics · Shuja Abidi

Pepper

Pepper: a humanoid built to talk, not to walk

Pepper has a head, a torso and two arms, but it rolls on wheels. It was made to meet people in shops, schools and labs, so its hard problems are about interaction rather than balance.

1

Introduction

A different kind of humanoid from the walking robots on the last page.

Pepper, a white humanoid robot with a tablet on its chest, on display
PepperSoftBank and Aldebaran · unveiled 2014Photo: Wikimedia Commons, CC0

Robots such as the Unitree G1 and Atlas are built to do physical work, so they must balance on two feet while they move. Pepper makes a different trade. Its wheeled base keeps it steady, and its design goes into the things people notice: a friendly shape, a head that can turn towards you, expressive arms and a tablet on its chest.

This makes it a social robot, one designed mainly to communicate with people. It has been widely used as a platform for human-robot interaction research, because a researcher can study how people respond to a robot without first having to solve walking.

Pepper isn't a stronger or weaker version of a walking humanoid. It answers a different question: not "how does a robot stay up?", but "how should a robot behave around people?"

2

The body

About 1.2 m tall, 28 kg, with 20 joints: 17 in the body and 3 in the wheels.

PartJointsWhat they do
Head2Turn and nod
Each arm5Shoulder pitch and roll, elbow yaw and roll, wrist yaw
Each hand1All the fingers open and close together
Hip2Lean forward, back and to the side
Knee1Bends the lower body, for example to lean or look down
Base3 wheelsOmnidirectional wheels: moves in any direction, or turns on the spot

Joint counts and limits come from the pepper_robot description. Some spec sheets give slightly different totals, depending on how they count the base.

What the numbers show

Gentle arms

The shoulder pitch joint is limited to about 5.4 N·m, and the hand motor to far less. The arms suit gestures and light touch, not lifting.

Short reach

Shoulder to wrist is about 33 cm, so the arms reach much less far than a person's.

A limit on shoulders

Shoulder roll only moves outward from straight ahead. The upper arm can't swing across the chest, which limits some human poses.

3

History

From a high-profile launch, through a production pause, to new owners.

  1. Unveiled.SoftBank Mobile and France's Aldebaran announce Pepper in Japan, as a robot that can read emotions and talk with people.
  2. Shops, offices and labs.Pepper goes on sale in Japan and later elsewhere, greeting customers and appearing in universities as a research robot.
  3. Production paused.Reports say SoftBank has paused making Pepper because of weak demand.
  4. Back to Aldebaran.Germany's United Robotics Group takes over SoftBank's European robotics business, which returns to the Aldebaran name.
  5. Insolvency and sale.Aldebaran enters insolvency. Maxvision of Shenzhen buys its core assets, including the rights to Pepper and its smaller sibling NAO.

Reports differ on some dates in 2021–2025. If you're planning to buy or use a Pepper, check current support with the seller first.

4

How it moves

With no legs to balance on, much of the walking humanoid's problem goes away.

Walking humanoid, such as the G1Pepper
Staying uprightAn active problem, solved many times a secondMostly handled by its wide wheeled base
Moving aroundLearned or planned footstepsA speed command to the wheels
Why RL is usedBalance and walking are hard to program by handNot needed for basic motion; useful for specific research tasks
Copying a personThe robot must also keep its balanceArm and head angles can usually be followed directly, within limits
Main challengeNot falling while doing the taskBehaving naturally and safely near people

How Pepper can copy a person's arms

  1. Find the body. A pose estimator such as MediaPipe finds the person's shoulders, elbows and wrists in a video.
  2. Find the arm directions. The direction of each upper arm and forearm is measured relative to the person's own chest, so it doesn't matter which way they stand.
  3. Solve the joint angles. Inverse kinematics finds the Pepper joint angles that point its arms the same way, within its joint limits.
  4. Slow it down. The motion is smoothed and speed-limited, so a noisy video frame can't become a sudden movement near someone.

Because the arm directions are copied, not the person's exact hand positions, a robot with a different body size can follow the same motion. Pepper's short arms and shoulder limits mean the copy will never be exact.

Does a wheeled robot never tip over?

It can. A wheeled base is stable only while the robot's centre of mass stays over the area between its wheels, and its motion doesn't push it outside. Leaning, fast arm motion, slopes and pushes all reduce that margin, so balance gets easier on wheels but doesn't disappear.

5

Open source or proprietary?

The robot's core software is closed, but there are open tools around it.

PartWhat it doesStatus
NAOqiThe robot's operating system and programming framework, also used by NAOProprietary
Choregraphe and Android SDKThe maker's tools for building Pepper applicationsProprietary
naoqi driver for ROSConnects the robot's sensors and motors to ROSOpen
pepper_robotThe robot's description for ROS, including joints and limitsApache-2.0
qiBulletA Python simulator for Pepper, NAO and Romeo. It downloads the robot meshes separatelyApache-2.0

Check the version before you start: some Python tools work with NAOqi 2.5 but not with NAOqi 2.9 or newer. Licences can also change, so check each repository before reuse.

6

What it suits

Choose it for interaction, not for physical work.

A good fit

  • Studies of how people respond to a robot
  • Conversation, gestures and gaze
  • Greeting, guiding and giving information
  • Light physical contact, such as a handshake
  • Teaching, because students can program it safely

A poor fit

  • Lifting or carrying anything heavy
  • Stairs and rough ground
  • Embracing a person: its arms are short and weak
  • Precise force control, because it's controlled mainly through joint positions and stiffness
  • Work that needs a long-term supply of new robots and parts