computation-driven design for legged robotsprof. stelian coros department of computer science...

The Computational Robotics Lab

Prof. Stelian Coros

Department of Computer Science

Computation-driven design for legged robots

Computational

Robotics Lab

The Computational Robotics Lab

The Computational Robotics Lab

The three pillars of our research program

dexterous

manipulation

form and function for

compliant systems

mobility

The Computational Robotics Lab

Robots need to become much easier to create, program and deploy

The Computational Robotics Lab

Robots need to become much easier to create, program and deploy

[Siggraph 2018]

The Computational Robotics Lab

Motion parameters:

Motion constraints:

n

t1

t2

Motion parameters:

Motion constraints:

𝜌(𝛼)𝜔(𝛼)

𝑡(𝛼)

𝑒

Motion parameters:

Motion constraints:

Motion parameters:

Stable, skilled motions are generated through optimization:

The Computational Robotics Lab [Siggraph 2018]

The Computational Robotics Lab

The Computational Robotics Lab 1

3

The Computational Robotics Lab

The Computational Robotics Lab

Forward and Inverse Design

• Forward design:

– User input: robot design

– System output: motion controller

Forward and Inverse Design

• Forward design:

– User input: robot design

– System output: motion controller

• Inverse design:

– User input: high-level task

– System output: robot design and motion controller

Specifying the morphological design for a robot that can accomplish a specific task can still be rather difficult

High-level task

Design Vocabulary

The Design Space

The Design Space

Heuristic-guided best-first search

Intermediate Design Hybrid Design

VirtualEnd Effector

Forward and Inverse Design

• Forward design:

– User input: robot design

– System output: motion controller

• Inverse design:

– User input: high-level task

– System output: robot design and motion controller

Motion parameters: p

Stable, skilled motions are generated through optimization:

, ∀𝐩

𝑑𝐦

𝑑𝐩= −

𝜕𝐆

𝜕𝐦

−𝟏 𝜕𝐆

𝜕𝐩

p

𝑑𝐆

𝑑𝐩=

𝜕𝐆

𝜕𝐩

𝑑𝐦

𝑑𝐩+ 𝜕𝐆

𝜕𝐩= 𝟎

𝚫𝐩 =𝜕𝐦

𝜕𝐩

+

𝚫𝐦

𝚫𝐩 =𝜕𝐦

𝜕𝐩

+

𝚫𝐦

𝚫𝐩 =𝜕𝐦

𝜕𝐩

+

𝚫𝐦

Actuation forces reduced by ~50%

||The Computational Robotics Lab

Where to next?

IHMC’s Nadia

Understand how robots can exploit soft

materials as effectively as living things do

||The Computational Robotics Lab

Thank you!

Sep 5, 2018Stelian Coros 38

Computational Robotics Lab, ETH Zurich