Transcript Slides

Randomized
KinoDynamic
Planning
Steven LaValle
James Kuffner
Randomized KinoDynamic Planning
• To determine the sequence of control inputs
to drive a robot from an initial state to an end
state while obeying physically based dynamic
models and avoiding obstacles in the robot’s
environment
• Approach: tailored form of randomization
(RRT’s) specially suited to high dimensional
state spaces
Problem Formulation
• Differential Constraints in State Space
Problem Formulation
• Handling Obstacles in State Space
Problem Formulation
• Solution Trajectory
Unique Challenges:
• State Space has twice the dimension of Configuration; worsens the curse
• Momentum considerations cause drift, overshooting, oscillations, which
potential fields and probabilistic roadmaps can’t handle
RRT based planner
• Rapidly exploring Random Trees- State Space
• Strategy…
– Initialize tree with vertex
– Repeatedly,
• select state at random in state space
• Select its nearest neighbor in the tree
• Choose a control and ensuing output that pushes
neighbor state towards random state
• Create new vertex and add to tree
RRT based planner
• Rapidly exploring Random Trees
• Outcome…
RRT based planner
• Rapidly exploring Random Trees
– PRO: works for high degrees of freedom
– PRO: no steering required
– PRO: biased toward unexplored space
– PRO: probabilistic completeness
– CON: no well defined metric
RRT based planner
• Bidirectionnal planning algorithm
• Strategy:
– Grow two RRT’s at initial and goal states
– At each growth step, check for intersection
– Either halt once path exists, or continue to
accumulate best paths
Spacecraft and Hovercraft tests
• Model
– State
Spacecraft and Hovercraft tests
• Model
– Control
Spacecraft and Hovercraft tests
• Model
– Metric (Euclidean)
– Weight vector is normalized
– Dot product represents cosine of angle
Spacecraft and Hovercraft tests
• Model
– Controls consists of a fixed set U in each example
– Each set includes a ‘no control’ control
– Each control is applied over a fixed timestep
• eg. dt = 0.01 sec
– Control timestep is independent of RRT timestep
Video
Spacecraft and Hovercraft tests
• Case 1: Planar Translating Body in X-Z plane
4 DOF: x,z,x’,z’
4 Controls:
Spacecraft and Hovercraft tests
• Case 2: Planar Body with Rotation
– 6 DoF: x,y,θ,x’,y’, θ’
– 3 controls:
• Translate forward
• Rotate clockwise
• Rotate counterCW
Spacecraft and Hovercraft tests
• Case 2: Planar Body with Rotation
– ~ 5 minutes
– 13,600 nodes
Spacecraft and Hovercraft tests
• Case 3: Translating 3-D body
– 6 DoF: x,y,z,x’,y’,z’
– 6 controls:
• Opposing forces In each of the 3 Principal directions
Spacecraft and Hovercraft tests
• Case 3: Translating 3-D body
– ~1 min
– 16,300 nodes
Spacecraft and Hovercraft tests
• Case 3: Translating 3-D body
– ~1 min
– 16,300 nodes
Spacecraft and Hovercraft tests
• Case 4: 3-D body with rotation
– Cylindrical satellite object
– 12 DoF: x,y,z,Rx,Ry,Rz
and derivatives
– 5 controls: translate along
Cylindirical axis, rotates arbitrarily
Simulates satellite docking
Spacecraft and Hovercraft tests
• Case 4: 3-D body with rotation
– ~6 minute
– 23,800 nodes
Spacecraft and Hovercraft tests
• Case 4: 3-D body with rotation
– ~6 minute
– 23,800 nodes
Spacecraft and Hovercraft tests
• Case 4: 3-D body with rotation
– 12 DoF
– 5 controls
– Forward, up-down
– Clockwise roll
Spacecraft and Hovercraft tests
• Case 4: 3-D body with rotation
– ~11 minute
– ??? nodes
Broader Applications
• Broad applications in Humanoid Robotics
– Generalized to many high DoF problems subject to
various constraints
• Integrated Grasp Planning (Vahrenkamp, Do et al) [6]
• Full Body Motion (S Kagami, J Kuffner et al) [7]