Documente Academic
Documente Profesional
Documente Cultură
1
Levels of DecisionMaking
§ Task: which object to move where?
§ Grasp planner: where to grasp object?
§ Inverse kinematics: arm/wholebody configuration
§ Motion planner: path through cspace
§ Trajectory following: joint torques/velocities
2
How to connect levels?
§ Standard model: topdown+heuristics
§ Decide which object to move next
§ Choose grasp point
§ Find IK solution
§ Etc
§ Protocol
§ High level gives low level a planning instance
§ Low level returns a single solution or fails 3
This talk
§ Angelic semantics: levels communicate via
(bounds on) reachable sets
§ Resulting planners
§ Do most work at high level => efficient
§ Yet provably sound and complete
4
Running example
O
O
1 Door 1 4
Robo
t
Goal
O2
Door 2
O5 O
3
O1
O4
Door 1
Robot
Goal
O2
Door 2
O5 O3 6
Planning problems
§ A (deterministic) planning problem consists of:
§ State space S
§ Goal set T ϵ S
§ Initial state s0 ϵ S
§ Primitive action set A
§ Transition function f : SxA → S
s
0
Transition
Transition function
function for
for action
action 7
aa12
Planning Problems
§ State § Actions
§ Robot configuration § Move (): directly
reachable
§ Confs of unheld objects
§ Grasp (o)
§ Held object (if any)
§ Ungrasp (o)
§ Relative conf of held object
§ Open (d)
§ Door statuses
O1 § Shut (d)
O4
Door 1
Robot
Goal
O2
Door 2
O5 O3
8
Hierarchies
§ A hierarchy for a planning problem consists of:
§ Set of highlevel actions H
§ Set of toplevel actions Top
§ For each high level action, set of refinements
§ Each refinement is a sequence of other actions
9
Hierarchy
Top Level
O1
O4
Door 1
Transfer (O1, G)
Transfer (O2, H) Robot
... Goal
O2
Door 2
O5 O3
G, H subsets of map
10
Hierarchy
Transfer (O4, G)
O1
O4
Door 1
Refinements
Robot
11
Hierarchy
GraspSide (O4, 2)
O1
O4
Door 1
Refinements
Robot
12
Hierarchy
Navigate (S)
O1
O4
Door 1
Refinements
Robot
Goal
NavToRoom(R2, Door1), Nav(S) O2
Door 2
NavToRoom(R2, Door2), Nav(S) O5 O3
13
Hierarchy
Nav (S)
O1
O4
Door 1
Refinements
Robot
Goal
Collisionfree kinematic motion plans O2
Door 2
O5 O3
ending in S
14
Refinements
Transfer (o2, H) Transfer (o1, G) Top Level Plan
…
Primitive
Move (1,2,.3) Move (2,3,.7) Grasp(o2) … Ungrasp (o1)Refinement
15
Hierarchical planning
§ Given
§ A planning problem
§ A hierarchy
§ Look for a plan that:
§ reaches the goal
§ is a primitive refinement of a toplevel plan
16
Using hierarchies to plan
§ # high level plans << # primitive plans
§ So we win if we can
§ Reject most bad plans at the high level
§ Often commit to good plans at high level
17
Using hierarchies to plan
O1
O4
Door 1
Robot
Goal
O2
Door 2
O5 O3
18
Using hierarchies to plan
§ # high level plans << # primitive plans
§ So we win if we can
§ Reject most bad plans at the high level
§ Often commit to good plans at high level
§ To do this we need transition models for HLAs
19
Our approach: Reachable sets
§ Reachable set of HLA h from state s:
{fa’(s) | a’ ϵ PrimRef(h)}
§ Angelic description of h Fh : S → 2S
Fh(s) = Reachable set of h from s
a1a2a
s h3
HLA: h3
Primitive refinements: a2a2
[a1a2a1], [a2a2]
20
Angelic nondeterminism
§ HLA descriptions map states to sets of states
§ May seem related to nondeterministic planning
§ Key difference: how actual successor state chosen
§ Nondeterministic planning: adversary chooses
§ Stochastic planning: nature chooses (by flipping a coin)
§ HLA descriptions: planner will eventually choose
21
Angelic descriptions…
1. Are true
3. Are composable
5. Support high level reasoning
22
Truth
§ Angelic HLA descriptions
§ Logically follow from primitive descriptions
§ Are empirically verifiable
§ Simplifies design/analysis of algorithms that
§ Derive HLA descriptions from primitive
§ Learn descriptions by observation
23
Composability
§ Can extend exact descriptions to:
§ Sets of states
§ Sequences of actions
s h h
0
h
1 2
1h2
h
2
24
High level reasoning
§ Given high level sequence p, the following
are equivalent
§ Fp(s0) intersects the goal set
§ p has a primitive refinement that succeeds
§ Allows rejecting/committing to plans at
high level
25
Optimistic descriptions
§ Exact description F could be complicated
§ NPHard to represent compactly
§ Optimistic description: superset of reachable set
s Fh(s
Gh(s
26
Optimistic descriptions
§ Transfer (o, G)
§ Object o must end up inside G
§ Robot must end up within distance .1 of G
§ NavToRoom (r)
§ If crossSection (robot + object) >= door width
§ Empty set (no reachable states)
§ Else
§ Robot must end up within room r
27
Properties of optimistic descriptions
§ Theorem: if the optimistic reachable set of plan
p does not intersect the goal, then no refinement
of p solves the problem
§ Allows pruning away plans at highlevel
28
Pessimistic descriptions
§ Dual to optimistic descriptions
§ Pessimistic description: subset of reachable set
s Fh(s
Gh(s
29
Pessimistic descriptions
§ Nav (G)
§ If world has no bottlenecks
§ Can reach any configuration in G
§ Else if exists a path in roadmap to set G’ in G
§ G’ achievable
§ Else
§ Empty set
30
Properties of pessimistic descriptions:
§ Theorem: if the pessimistic reachable set of plan
p intersects the goal, then some primitive
refinement of p solves the problem
§ Allows committing to plans at highlevel
31
Numerical costs
§ So far, we’ve restricted to goalbased problems
§ What if actions have numerical costs/rewards?
§ Generalize descriptions, so HLA has
§ Pessimistic cost – upper bound on achievable cost
§ Optimistic cost – lower bound on achievable cost
§ Example: Nav(G)
§ Optimistic cost: at least Euclidean distance to G
§ Pessimistic cost: distance in roadmap
32
Planning algorithms
§ Repeat
§ Get next plan
§ If fails optimistically, reject
§ If succeeds pessimistically, commit
§ Else, add refinements to set
§ Examples
§ AHA*: find hierarchically optimal plan
§ SHP: find any plan below a cost threshold 33
Experiments
§ Discretized manipulation problem
§ Easier instance: 42 step plan required
34
Experiments
§ Harder instance
§ 284 step plan required
§ Nonhierarchical methods did not terminate within 10,000 seconds
35
Online navigation planning
36
Online navigation planning
37
Conclusions
§ Angelic hierarchical planning
§ Reason about reachable sets of HLAs
§ Reject/commit to plans at high level when possible
§ Descend to lower levels when necessary
§ Future work
§ Extending planners to return reachable sets
§ Learning hierarchies and descriptions
§ Probabilistic dynamics
Thank you! 38