Declarative Programming for Modular Robots Ashley-Rollman, De Rosa, Srinivasa, Pillai, Goldstein, Campbell

1
Declarative Programming for Modular
RobotsAshley-Rollman, De Rosa, Srinivasa,
Pillai, Goldstein, Campbell

• November 2, 2007

2
Locally Distributed Predicates (LDP) Meld

• Two very different approaches to declarative
  programming for modular robots
• Meld - logic programming
• LDP - distributed pattern matching
• Both achieve higher goals
• Dramatically shorter code
• Automatically distributed
• Automatic messaging

3
LDP Overview

• Originated in Distributed Watchpoint system
• Needed to describe and detect incorrect
  distributed state configurations
• Locally Distributed Predicates
• Locally Distributed involving a bounded number
  of connected modules
• Predicates boolean expressions over state,
  temporal, and topological variables
• An LDP program consists of a number of
  predicates, each with one or more attached
  actions
• Every predicate/action pair is executed in
  parallel

4
Meld Overview

• Logic programming language
• Inspired by P2 Loo et. al. 2005
• Consists of facts and rules for deriving new
  facts
• When a fact changes, derived facts are
  automatically deleted
• Programs typically consider local neighborhoods
• Additional support for making non-local
  neighborhoods

5
LDP and Meld A Comparison
LDP Meld
Approach Predicate Matching Proof Search
Concise Code (vs. C/Java) ? ?
Automatic Messaging ? ?
Operations over all neighbors count forall
Quantified (Variably-Sized) Expressions ?
Automatic Fact Retraction ?
State Management By Programmer By System
6
Example 3D Shape Change Algorithm

• lt20 lines of Meld or LDP
• Connectivity maintenance guaranteed by algorithm

7
Example 1 Setting Module State
type state(module, min int). state(A, FINAL) -
isSeed(A). state(B, FINAL) - neighbor(A,
B), state(A, FINAL), in(B).
forall (a) where (a.isSeed) do a.state
FINAL forall (a,b) where (a.state FINAL)
(b.inside) do b.state FINAL
LDP
Meld

• If the module is the seed
• Set the seeds state to FINAL
• For every module inside the target shape
• If it is next to a module in FINAL state
• Set the modules state to FINAL

8
LDP and Meld A Comparison
LDP Meld
Approach Predicate Matching Proof Search
Concise Code (vs. C/Java) ? ?
Automatic Messaging ? ?
Operations over all neighbors count forall
Quantified (Variably-Sized) Expressions ?
Automatic Fact Retraction ?
State Management By Programmer By System
9
Example 2 Evaluation Over all Neighbors
type deletable(module). type notChild(module,
module). notChild(A, B) - neighbor(A, B),
parent(B, C), A ! C. deletable(A) -
forall neighbor(A, B) notChild(A, B).
forall(a,b) where (b.parent ! a.id) do
a.notChild.add(b.id) forall(a) where
size(a.notChild) size(a.neighbors) do
a.delete()
LDP
Meld

• A module can only be deleted if none of its
  neighbors are children
• We first determine which neighbors are not
  children
• If there are no children, the module can be
  deleted

10
LDP and Meld A Comparison
LDP Meld
Approach Predicate Matching Proof Search
Concise Code (vs. C/Java) ? ?
Automatic Messaging ? ?
Operations over all neighbors count forall
Quantified (Variably-Sized) Expressions ?
Automatic Fact Retraction ?
State Management By Programmer By System
11
Example 3 Self-deleting Gradients
12
Example 3 Self-deleting Gradients
forall (a,b) where (a.value gt b.value) do
a.value b.value 1 forall (a,b0,6) where
count(a.value gt bi.value) 0 a.value ! 0
do a.value INF
type gradient (module, min int). gradient(A, N)
- neighbor(A, B),
gradient(B, M), N M
1.
LDP
Meld

• Meld deletes all dependent facts when the root
  fact is deleted
• LDP directly manipulates state variables, so
  retraction must be manual
• LDP must specify the maximum number of neighbors

13
LDP and Meld A Comparison
LDP Meld
Approach Predicate Matching Proof Search
Concise Code (vs. C/Java) ? ?
Automatic Messaging ? ?
Operations over all neighbors count forall
Quantified (Variably-Sized) Expressions ?
Automatic Fact Retraction ?
State Management By Programmer By System
14
Example 4 Spanning Tree Creation
15
Example 4 Spanning Tree Creation
forall (a) where (a.isRoot 1) do a.parent
a.id forall (a,b) where (a.parent ! -1)
(b.parent -1) do b.parent a.id
type parent(module, first module). parent(A, A)
- root(A). parent(B, A) - neighbor(B, A),
parent(A, _).

• Newer versions of Meld use the first aggregate
  to ensure uniqueness
• This qualifier is not sufficient for more complex
  situations

16
Example 4b Spanning Tree Creation
forall (a) where (a.isRoot 1) do a.parent
a.id forall (a,b) where (a.parent ! -1)
(b.parent -1) do b.parent a.id
type possibleParent(module, module, int). type
bestParent(module, min int). type parent(module,
module). parent(A, A) - root(A). possibleParent(B
, A, T) - neighbor(A, B), parent(A,
_) , T
localTimeStamp(). bestParent(B, T) -
possibleParent(B, _, T). parent(B, A) -
possibleParent(B, A, T), bestParent(B,
T).

• Without first, Meld must use timestamps to
  ensure exactly one unique parent
• LDP uses a single state variable, and thus can
  never have more than one parent

17
LDP and Meld A Comparison
LDP Meld
Approach Predicate Matching Proof Search
Concise Code (vs. C/Java) ? ?
Automatic Messaging ? ?
Operations over all neighbors count forall
Quantified (Variably-Sized) Expressions ?
Automatic Fact Retraction ?
State Management By Programmer By System
18
Future Research

• Performance enhancements/optimizations
• Additional language features
• Support transactions
• Applicability to other application domains
• Explore tradeoffs between automated and manual
  state control
• Find a balance that allows programmers to
  maintain state while gaining some or all of the
  benefits of automated state

Interested in Meld/LDP? Email mderosa,mpa_at_cs.cmu
.edu