• No results found

Safe robot motion planning in dynamic, uncertain environments

N/A
N/A
Protected

Academic year: 2021

Share "Safe robot motion planning in dynamic, uncertain environments"

Copied!
23
0
0

Loading.... (view fulltext now)

Full text

(1)

Safe robot motion planning in dynamic, uncertain environments

RSS 2011 Workshop: Guaranteeing Motion Safety for Robots

June 27, 2011

Noel du Toit and Joel Burdick California Institute of Technology

(2)

Dynamic, Cluttered, Uncertain Env’s

6/27/2011 Noel du Toit 2

(3)

Conceptual Problem

Components:

Localization Process Noise Mapping

Detection/

tracking Prediction

Characterization Occlusions/

drop-outs

Problem

Formulation?

Computation?

(4)

Probabilistic uncertainty => conservative

Conservatism: incorporate anticipated measurements

Previous works: static environments [Prentice 09], [Censi 08], [Yan 05]

Enabling capability: complicated agent behaviors, more clutter, agent information gathering

Safety: chance constraint conditioning

Safety vs Conservatism

6/27/2011 Noel du Toit 4

(5)

History I-state:

Belief state: probability distribution

State:

Transition function:

Cost Function

Encodes noise properties &

and dynamics

Stochastic Dynamic Programming (SDP)

state control

measurement noise process noise

 

 

 

 

=

An

i A i

R i

i

x x x

x

1

Terminal cost

Stage cost

(6)

6/27/2011 Noel du Toit 6

Control Policy:

Problem:

Stochastic Dynamic Programming (SDP)

Stochastic Dynamic Programming (SDP)

Feedback over all possible future measurements and the resulting belief states

(7)

Stochastic Dynamic Programming (SDP)

Issues:

Computationally intensive [Thrun et al. 05], [Bertsekas 05]

No hard constraints

Solution:

for Linear, quadratic cost, Gaussian noise [Bertsekas 05], [Bar-Shalom 81]

Approximations

Value & policy iteration [Thrun et al. 05],

Roll-out algorithm [Bertsekas 05]

Restricted structure approximations [Bertsekas 05]

(8)

Formulation: Stochastic RHC (SRHC)

Stochastic system:

Expected cost, chance constraints

Belief state:

Transition function:

Disturbance:

6/27/2011 Noel du Toit 8

(9)

Approximation: Most Likely Measurement

Most likely measurement:

Same computational benefits of OLRHC approach

Approximation: relies on RHC formulation to correct for assumed information to reduce conservatism in solutions

Effect of measurements

Update covariance

Update mean

Theorem [duToit ‘09] : The Most Likely Measurement Approx.

introduces no new information Least Informative Approximation

(10)

Chance Constraints

Constrain uncertain state

Probabilistic Collision Avoidance

Robot and obstacle uncertainty

Joint distribution and indicator function

6/27/2011 Noel du Toit 11

(11)

Collision Constraints: Evaluation

Monte-Carlo Simulation

Small level of confidence: requires ~10000 samples

~5ms per evaluation

Approximate: small disk/ellipse objects

Independent, Gaussian distributions:

Dependent, Gaussian distributions

6/27/2011 Noel du Toit 12

(12)

Safety: Reaction Horizon

Quantify time (# of stages) to react to changes in environment

Robot dynamics

Environment

Reaction horizon,

r

: react to modeled disturbances

Chance constraint conditioning:

Use

r

-stage open-loop predicted distribution

Anticipated information:

leverage PCL reduction in conservatism

6/27/2011 Noel du Toit 13

(13)

Safety vs Conservatism

Uncertainty grows over reaction horizon

Next stage: new information + re-solve problem (RHC)

PCL: leverage new information

OL: uncertainty growth results in conservative solutions

(14)

1-D example:

6/27/2011 Noel du Toit 15

Car following:

Collision constraint (maintain some separation distance)

Velocity controlled random-walk model

Reaction horizon:

r=2

(to influence position)

(15)

1-D Example (cont’d)

Reaction horizon = 1

Reaction horizon = 2

Plot separation distance

(16)

Dynamic Environment: Oncoming Agents

6/27/2011 Noel du Toit 17

OLRHC PCLRHC

(17)

Summary

Practical systems: trade off conservatism and safety

PCLRHC

Reduce conservatism through anticipated information

RHC: resolve problem to incorporate actual measurements

Chance constraint conditioning

Allow for modeled disturbances

Can still leverage anticipated information

See Noel’s thesis for various variations on this problem

(18)

Thank you

[email protected]

[email protected]

Questions?

Publications:

Du Toit, N.E. and Burdick, J.W., “Robot Motion Planning in Dynamic, Cluttered, Uncertain Environments”, accepted to IEEE Transactions on Robotics

Du Toit, N.E. and Burdick, J.W., “Probabilistic Collision Checking with Chance Constraints”, accepted to IEEE Transaction on Robotics

Du Toit, N.E., “Robot Motion Planning in Dynamic, Cluttered, Uncertain Environments: the Partially Closed-Loop Receding Horizon Control Approach”, Ph.D. thesis, Caltech, 2010

Conferences:

Workshop on Motion Planning: From Theory to Practice (RSS) 2010

Du Toit, N.E. and Burdick, J.W., “Robotic Motion Planning in Dynamic, Cluttered, Uncertain Environments”, ICRA 2010

6/27/2011 Noel du Toit 19

(19)

Robot:

Agent:

Constraints:

Objective function:

Problem Definition

(20)

Related Work

Deterministic Probabilistic

Static Dynamic Static Dynamic

Control OC [Friedland 05]

RHC [Mayne 00]

OC with augmented states

OC with separation [Friedland 05]

Stochastic RHC (later)

Robotics Graph search, roadmap methods, etc. [LaValle 06], [Choset et al. 07]

Dynamic window [Fox et al. 97]

Velocity obstacles [Fiorini & Shiller 98]

Graph search in extended state space [LaValle 06], [Censi 08]

Probabilistic velocity obstacles [Fulgenzi et al. 07]

AI

MDPs [LaValle 06] Extended state space (time x pose) [LaValle 06]

DP [Bertsekas 05]

MDPs, POMDPs [Thrun et al. 05]

6/27/2011 Noel du Toit 21

Stochastic systems: Probabilistic vs. non-deterministic

AI: Artificial Intelligence OC: Optimal Control

RHC: Receding Horizon Control

DP: Dynamic Programming MDP: Markov Decision Process POMDP: Partially Observable MDP

PCLRHC

(21)

Most likely measurement:

Restricted information:

Resulting belief state:

Deterministic in belief state:

Partially Closed-loop RHC

(22)

Robot: Linear model, linear measurements

Velocity constraints:

Control constraints:

Collision constraints:

Agent: Linear constant velocity model, linear measurements

Simulation Setup

6/27/2011 Noel du Toit 23

(23)

PCLRHC Approximation

Information gain: relative entropy

Baseline:

PCL distribution:

Executed distribution:

References

Related documents

BMIDECR1 Bus Master IDE Command Register 1 for Second Channel (R/W) PCI Configuration Space 78h or Base Address #4 + 08h Bit 7 Reserved 6 Reserved 5 Reserved 4 Reserved. 3 Read or

However the researcher found there are no significant differences at α ≤ 0.05 between the means of organizational factors affecting professional nurses performance at north

u copper impedes sperm transport and viability in the cervical

In particular, the aims of the present study were to examine changes and age- related differences on the mobility performance with an additional cognitive or motor task, and to

research objectives are four-fold: 1) to examine visitors’ experiences, satisfaction, and revisit intention, 2) to investigate what specific dimension of experience will

Therefore, the aim of this study was to explore and identify medicinal plant species used by diabetic patients in Region of Fez-Meknes.. Materials and methods

To generate the edges in the network, we measured the semantic similarity of each pair of pathway nodes based on their associated GO terms in the minimized functional annotation pro