Legged Locomotion Planning Kang Zhao B659 Intelligent Robotics Spring 2013 1.

Post on 17-Jan-2016

220 views 1 download

Tags:

Transcript of Legged Locomotion Planning Kang Zhao B659 Intelligent Robotics Spring 2013 1.

1

Legged Locomotion Planning

Kang ZhaoB659 Intelligent RoboticsSpring 2013

2

Planning Biped Navigation Strategies in Complex Environments• Joel Chestnutt, James

Kuffner, Koichi Nishiwaki, Satoshi Kagami

3

O Global terrain map MO GoalO Primitive set {Trans}O Search algorithm

4

Algorithm - Biped Robot ModelO State:

O θ: position and orientation relative to {U}

O One-step motion destination

5

Algorithm- State transitionsO Footstep transition

0

1

2 34

5 6

7

A 16-transitions set

Branching factor

6

Algorithm- EnvironmentO Terrain map

7

Algorithm- State Evaluation

𝑉= 𝑓 (𝑄 ,𝑇 ,𝑄𝑐 ,𝑄𝑔)Location metric to

evaluate a location’s cost

𝐿𝑖={ 𝐿𝑖 (𝑄 )∞ 𝑖𝑓 𝐿𝑖 (𝑄 )>𝐿𝑖

𝑙𝑖𝑚𝑖𝑡, 𝑖=1…5

𝐿 (𝑄 )=∑𝑤𝑖𝐿𝑖

Slope angle

Roughness

Stability

Largest bump

Safety

8

Slope angle

Roughness

Stability

Largest bump

Safety

The slope angle of the surface at the candidate location. Perfectly horizontal surfaces are desired. The slope angle is computed by fitting a plane h(x, y) to the cells in the location.

1𝑁 ∑

𝑐∈𝐶

¿𝑐 . h h𝑒𝑖𝑔 𝑡−h (𝑐 . 𝑥 ,𝑐 . 𝑦 )∨¿¿

max {𝑐 . h h𝑒𝑖𝑔 𝑡−h (𝑐 . 𝑥 ,𝑐 . 𝑦 )∨𝑐∈𝐶 }

It’s purpose is to take into account the possible inaccuracy of foot positioning. This can be computed using the roughness and largest bump metrics, using the cells around the foot location

9

Algorithm- State Evaluation

Step metric to evaluate cost

of taking a step

𝑆 (𝑄 ,𝑇 ,𝑄𝑐 )=𝑇 .𝑐𝑜𝑠𝑡+𝑤h∨𝐻 (𝑄 ,𝑄𝑐 )∨¿

Cost of transition

• Penalty for height change• Collision check

𝑄𝑐=𝑇 (𝑄 )𝐻={ 𝐻 (𝑄 ,𝑄𝑐 )

∞ 𝑖𝑓 𝐻 (𝑄 ,𝑄𝑐 )>𝐻❑𝑙𝑖𝑚𝑖𝑡

𝑉= 𝑓 (𝑄 ,𝑇 ,𝑄𝑐 ,𝑄𝑔)

10

Algorithm- State Evaluation

Heuristic metric to evaluate

remaining cost

𝑅 (𝑄 ,𝑄𝑔)=𝑤𝑑𝐷 (𝑄 ,𝑄𝑔 )+𝑤𝜃|Ɵ (𝑄 ,𝑄𝑔 )|+𝑤h∨𝐻 (𝑄 ,𝑄𝑔 )∨¿

Euclidean distance Relative angle Height

difference

𝑉= 𝑓 (𝑄 ,𝑇 ,𝑄𝑐 ,𝑄𝑔) The heuristic function estimates the cost to go from to a goal state

Its value is independent of the current search tree; it depends only on and the goal

11

Best First SearchO It exploits state description to estimate how

“good” each search node isO An evaluation function maps each node of

the search tree to a real number

O Greedy BFS

𝑅 (𝑄 ,𝑄𝑔)

h (𝑁 )=𝑅 (𝑄 ,𝑄𝑔)

12

A* Search

h (𝑁 )=𝑅 (𝑄 ,𝑄𝑔)

𝑅 (𝑄 ,𝑄𝑔)

𝐿 (𝑄𝑐 )+∑ 𝑆 (𝑄𝑖 ,𝑇 ,𝑄𝑐)

Search tree

Searching the State SpaceA schematic view

Q s

Q g

Search tree

Searching the State SpaceA schematic view

Q s

Q g

T 1

T 2

Search tree

Searching the State SpaceA schematic view

Q s

Q g

Search tree

Searching the State SpaceA schematic view

Q s

Q g

Search tree

Searching the State SpaceA schematic view

Q s

Q g

Search tree

Searching the State SpaceA schematic view

Q s

Q g

19

ResultsO Cluttered terrain

20

ResultsO Multi-level terrain

21

ResultsO Uneven ground with obstacles

22

Comparisons

O Distance to goalO Transitions and obstacle effectsO Metric weights

23A 26-transitions set

A 40-transitions set BFS

24

Performance comparison of A* and BFS for increasing numbers of stairs along the path

25

26

𝑅 (𝑄 ,𝑄𝑔)=𝑤𝑑𝐷 (𝑄 ,𝑄𝑔 )+𝑤𝜃|Ɵ (𝑄 ,𝑄𝑔 )|+𝑤h∨𝐻 (𝑄 ,𝑄𝑔 )∨¿

27

Local-minimum problem

28

Online Experiments

Stereo vision system

PlannerFootstep sequence

Trajectory generator

Walking area map

29

Following work

O A tired planning Strategy for biped navigation, 2004O Biped navigation in rough environments using

on-board sensing, 2009

30

Multi-Step Motion Planning for Free-climbing Robots• Tim Bretl, Sanjay Lall,

Jean-Claude Latombe, Stephen Rock