EASy 5 Dec 2004HART 2004, Fukui1 HART 2004 Time and Motion Studies: The Dynamics of Cognition, Computation and Humanoid Walking Inman Harvey, Eric Vaughan,

5 Dec 2004HART 2004, Fukui 1

EASy

HART 2004HART 2004HART 2004HART 2004

Time and Motion Studies:

The Dynamics of Cognition,

Computation and Humanoid WalkingInman Harvey, Eric Vaughan, Ezequiel Di Paolo

Evolutionary and Adaptive Systems Group

EASy, Dept. of Informatics

University of Sussex

[email protected]

5 Dec 2004HART 2004, Fukui 2

EASy

5 Dec 2004HART 2004, Fukui 3

EASy

HOAP-2, FujitsuHOAP-2, FujitsuHOAP-2, FujitsuHOAP-2, Fujitsu

5 Dec 2004HART 2004, Fukui 4

EASy

What are the difficulties?What are the difficulties?What are the difficulties?What are the difficulties?

Why is walking so easy for us, and so difficult for robots?

We suggest:-

Commercial humanoid design has tended to ignore the natural dynamics of a mechanical system, for at least two reasons.

(1) The conceptual framework of traditional A.I.(2) The background of industrial robot designers

5 Dec 2004HART 2004, Fukui 5

EASy

Instead, we should be … …Instead, we should be … …Instead, we should be … …Instead, we should be … …

… … exploiting the natural dynamics of mechanical limbs in ways comparable to those of animals and humans.

We should be producing designs that are not constrained by the shackles of GIFAI (Good Old Fashioned AI), nor by the constraints of conventional engineering design philosophy.

The Dynamical Systems approach, with Evolutionary Robotics

5 Dec 2004HART 2004, Fukui 6

EASy

The Glider Analogy …The Glider Analogy …The Glider Analogy …The Glider Analogy …

Compare the problems of designing humanoid walking robots with the problems that the Wright brothers faced in aiming towards the first powered flight on 17th Dec 1903.

They started by understanding un-powered gliding flight, before then adding control and then power.

5 Dec 2004HART 2004, Fukui 7

EASy

… … applied to humanoid walkingapplied to humanoid walking… … applied to humanoid walkingapplied to humanoid walking

Passive Dynamic Walking is the equivalent of the glider

Then we can progressively add Control and Power

5 Dec 2004HART 2004, Fukui 8

EASy

Plan of TalkPlan of TalkPlan of TalkPlan of Talk

1. Historical roots of Humanoid robotics

2. Models of Cognition – GOFAI

3. Models of Cognition – Dynamical Systems approach

4. Designing Dynamical Systems – Evolutionary Robotics

5. GOFAI Humanoid walking – ZMP methods

6. DS Humanoid Walking – PDW Passive Dynamic Walkers

7. Adding Control and Power to PDW

5 Dec 2004HART 2004, Fukui 9

EASy

Automata have a rich historyAutomata have a rich historyAutomata have a rich historyAutomata have a rich history

Hero of Alexandria described working models of animals and humans, using hydraulics and pneumatics, some 2000 years ago

5 Dec 2004HART 2004, Fukui 10

EASy

Clockwork technologyClockwork technologyClockwork technologyClockwork technology

From the 14th Century on, clockwork allowed more sophisticated automata

18th C, Jaquet-Droz

5 Dec 2004HART 2004, Fukui 11

EASy

Clockwork computingClockwork computingClockwork computingClockwork computing

In the 1820s in London, Babbage used clockwork technology to design the Difference Engine, and then the Analytical Engine

… … the world’s first universal digital computer

5 Dec 2004HART 2004, Fukui 12

EASy

KarakuriKarakuriKarakuriKarakuri

Hanzo Yorinao Hosokawa was a Master of making mechanical puppets in the 18th C.

5 Dec 2004HART 2004, Fukui 13

EASy

Tanaka HisashigeTanaka HisashigeTanaka HisashigeTanaka Hisashige

Established a Hall of Automata in Kyoto in the 19th C.

(child with bow and arrow)

He went on to build he first steam locomotive in Japan, and contributed to the industrialisation of Japan.

5 Dec 2004HART 2004, Fukui 14

EASy

From Karakuri to AsimoFrom Karakuri to AsimoFrom Karakuri to AsimoFrom Karakuri to Asimo

Walking karakuri inspired humanoid robots, especially in Japan.

From Waseda University in the 1960s, to Honda’s Asimo in the 1990s.

5 Dec 2004HART 2004, Fukui 15

EASy

Plan of TalkPlan of TalkPlan of TalkPlan of Talk

1. Historical roots of Humanoid robotics

2. Models of Cognition – GOFAI

3. Models of Cognition – Dynamical Systems approach

4. Designing Dynamical Systems – Evolutionary Robotics

5. GOFAI Humanoid walking – ZMP methods

6. DS Humanoid Walking – PDW Passive Dynamic Walkers

7. Adding Control and Power to PDW

5 Dec 2004HART 2004, Fukui 16

EASy

Constraints arising from this historyConstraints arising from this historyConstraints arising from this historyConstraints arising from this history

The design of such humanoids has been heavily influenced by this history.

1. It has been assumed that the trajectories of limbs must be pre-planned through computations

2. Designers have tended to use stiff actuators and materials, so as to maintain close control on these pre-calculated positions

5 Dec 2004HART 2004, Fukui 17

EASy

Computing and GOFAIComputing and GOFAIComputing and GOFAIComputing and GOFAI

Babbage’s computers in the 19th C were based on clockwork, and 20th C computers as invented by Turing and Von Neumann proceed by the ticks of a clock.

The dynamics of time, the sun, and a sundial are continuous

… but a clock and a computer go in a discrete sequence of ticks

5 Dec 2004HART 2004, Fukui 18

EASy

GOFAI assumptionsGOFAI assumptionsGOFAI assumptionsGOFAI assumptions

So there is a natural tendency to analyse even a dynamical process such as biped walking as a succession of moves between instantaneous frozen positions.

The dynamic has been reduced to transitions between static snapshots.

This is one cause of the rather unnatural underlying principles of many commercial bipeds today.

5 Dec 2004HART 2004, Fukui 19

EASy

Plan of TalkPlan of TalkPlan of TalkPlan of Talk

1. Historical roots of Humanoid robotics

2. Models of Cognition – GOFAI

3. Models of Cognition – Dynamical Systems approach

4. Designing Dynamical Systems – Evolutionary Robotics

5. GOFAI Humanoid walking – ZMP methods

6. DS Humanoid Walking – PDW Passive Dynamic Walkers

7. Adding Control and Power to PDW

5 Dec 2004HART 2004, Fukui 20

EASy

The Dynamical Systems approachThe Dynamical Systems approachThe Dynamical Systems approachThe Dynamical Systems approach

In contrast to GOFAI:-

The limbs of an animal, a human, or a robot – and their nervous systems, real or artificial – are physical systems with positions and values acting on each other smoothly in continuous real time.

Walking has a natural dynamics arising from the swing of limbs under gravity – synthesis an artificial system that respects such natural dynamics

5 Dec 2004HART 2004, Fukui 21

EASy

Passive Dynamic WalkingPassive Dynamic WalkingPassive Dynamic WalkingPassive Dynamic Walking

With upper and lower legs, and un-powered thigh and knee joints, a biped can walk down a slope with no control system

… in simulation …

5 Dec 2004HART 2004, Fukui 22

EASy

… … or in Realityor in Reality… … or in Realityor in Reality

Collins,

Cornell.

5 Dec 2004HART 2004, Fukui 23

EASy

Plan of TalkPlan of TalkPlan of TalkPlan of Talk

1. Historical roots of Humanoid robotics

2. Models of Cognition – GOFAI

3. Models of Cognition – Dynamical Systems approach

4. Designing Dynamical Systems – Evolutionary Robotics

5. GOFAI Humanoid walking – ZMP methods

6. DS Humanoid Walking – PDW Passive Dynamic Walkers

7. Adding Control and Power to PDW

5 Dec 2004HART 2004, Fukui 24

EASy

Evolutionary RoboticsEvolutionary RoboticsEvolutionary RoboticsEvolutionary Robotics

Rather than forcing limbs to follow a pre-planned trajectory, we want to design in the appropriate natural dynamics.

When we add an artificial nervous system, we also want to design in the appropriate dynamics for this, coupled to actuators and sensors.

We are designing dynamical systems, not calculating trajectories – this can be difficult.

Animals and humans are designed through Darwinian evolution – we can use Artificial Evolution

5 Dec 2004HART 2004, Fukui 25

EASy

Artificial EvolutionArtificial EvolutionArtificial EvolutionArtificial Evolution

1. Set up a mapping from strings of “Artificial DNA” to designs of robot bodies and their “nervous systems”

2. Start with a random population of DNA-strings – generating random and probably useless designs

3. Test each design, and pick out the “fitter” ones

4. Breed from the fitter ones – recombine and mutate their DNA to produce offspring

5. This makes a new generation – return to (2)

5 Dec 2004HART 2004, Fukui 26

EASy

Evolutionary RoboticsEvolutionary RoboticsEvolutionary RoboticsEvolutionary Robotics

The DNA will encode

1. The lengths, centres of mass, angles, spring constants, range of motor forces of the robot body

2. The connectivities, weights, biases and time parameters of a robot nervous systems

5 Dec 2004HART 2004, Fukui 27

EASy

TestingTestingTestingTesting

Typically the designs are tested in a physics simulator such as ODE.

Noise or uncertainty is added to the dimensions of the robot, to the physical forces – and where there is a control system, to sensors, actuators and the “nervous system”.

The simulations are computational – but the designs being simulated are not. They are real-time Dynamical Systems.

An “Envelope of Noise” can assist in transferring from simulation to a real physical robot (Jakobi et al)

5 Dec 2004HART 2004, Fukui 28

EASy

Plan of TalkPlan of TalkPlan of TalkPlan of Talk

1. Historical roots of Humanoid robotics

2. Models of Cognition – GOFAI

3. Models of Cognition – Dynamical Systems approach

4. Designing Dynamical Systems – Evolutionary Robotics

5. GOFAI Humanoid walking – ZMP methods

6. DS Humanoid Walking – PDW Passive Dynamic Walkers

7. Adding Control and Power to PDW

5 Dec 2004HART 2004, Fukui 29

EASy

GOFAI Humanoid DesignGOFAI Humanoid DesignGOFAI Humanoid DesignGOFAI Humanoid Design

Because of the historical constraints on their approach, GOFAI designers have used stiff actuators and materials.

High impedance systems, so that unplanned variations are resisted.

The Honda robot uses a version of ZMP, Zero-Moment Point Control, that requires it to accurately obey precisely calculated trajectories – only modified by force sensors in the ankles – a high impedance solution.

5 Dec 2004HART 2004, Fukui 30

EASy

ZMPZMPZMPZMP

Engineers find it much easier to measure positions rather than forces. Hence the tendency towards stiff, high-impedance solutions.

If the leg of a biped is allowed to straighten up at the knee, then ZMP calculations have a singularity with no sensible solution – hence the characteristic bent knees.

5 Dec 2004HART 2004, Fukui 31

EASy

Plan of TalkPlan of TalkPlan of TalkPlan of Talk

1. Historical roots of Humanoid robotics

2. Models of Cognition – GOFAI

3. Models of Cognition – Dynamical Systems approach

4. Designing Dynamical Systems – Evolutionary Robotics

5. GOFAI Humanoid walking – ZMP methods

6. DS Humanoid Walking – PDW Passive Dynamic Walkers

7. Adding Control and Power to PDW

5 Dec 2004HART 2004, Fukui 32

EASy

Remember the Glider AnalogyRemember the Glider AnalogyRemember the Glider AnalogyRemember the Glider Analogy

Our approach, following the Wright brothers, is to perfect a glider first, then add power and control.

Can the PDW scale up? How about with 4 knees?

5 Dec 2004HART 2004, Fukui 33

EASy

10 degrees of freedom version10 degrees of freedom version10 degrees of freedom version10 degrees of freedom version

Damped ankle and hip springs

5 Dec 2004HART 2004, Fukui 34

EASy

Passive Dynamic WalkingPassive Dynamic WalkingPassive Dynamic WalkingPassive Dynamic Walking

No power except potential energy, no control system

5 Dec 2004HART 2004, Fukui 35

EASy

Plan of TalkPlan of TalkPlan of TalkPlan of Talk

1. Historical roots of Humanoid robotics

2. Models of Cognition – GOFAI

3. Models of Cognition – Dynamical Systems approach

4. Designing Dynamical Systems – Evolutionary Robotics

5. GOFAI Humanoid walking – ZMP methods

6. DS Humanoid Walking – PDW Passive Dynamic Walkers

7. Adding Control and Power to PDW

5 Dec 2004HART 2004, Fukui 36

EASy

Adding Control and PowerAdding Control and PowerAdding Control and PowerAdding Control and Power

In the Passive versions, there are no sensors, no motors no control system.

Add force sensors, accelerometers, gyroscopes, rotation sensors.

Add actuators connected to springs – low impedance.

Add continuous time recurrent neural networks – and evolve them.

5 Dec 2004HART 2004, Fukui 37

EASy

Some examplesSome examplesSome examplesSome examples

Powered walking on the flat

Balancing on a moving platform

Walking back and forward (2D version)

5 Dec 2004HART 2004, Fukui 38

EASy

ConclusionsConclusionsConclusionsConclusions