Humanlike Avatar Navigation in Complex Environments

1
Human-like Avatar Navigation in Complex
Environments

• Sachin Patil

2
Avatar Navigation in Complex Environments

• Adapted from A Grasp Based Motion Planning
  Algorithm for Character Animation M. Kalisiak,
  M.V. Panne, EG00

3
Why is this important?

• Simulating Virtual Reality scenarios (urban,
  military, industrial) and virtual prototyping
• SecondLife is a classical example
• Avatar animation in the entertainment industry
• Important from a robotics perspective as lots of
  recent research is focused on highly articulated
  humanoid robots

4
Why is this difficult?

• Human beings have a very high number of DOFs
  (120) which makes planning using traditional
  techniques very difficult
• Human beings have kinodynamic constraints, which
  are difficult to formulate
• Most applications demand interactive real-time
  animation
• Most VR simulation scenarios (or even the real
  world for that matter) are dynamic in nature

5
Why is this difficult?

• Lastly, human avatar animation lies in the realm
  of the Uncanny Valley

( Source Wikipedia )
6
Limitations of standard motion planning approaches

• C-space complexity blows up
• Paths returned by randomized planners are jagged
  and unnatural
• Quality of human motion is easily discernable by
  the naked eye
• Kinodynamic planning, especially in dynamic
  environments, is a very new, active area of
  research in motion planning
• Extending it to high DOF robots is non-trivial

7
Two ways to tackle the problem

• Treat the navigation problem as one high
  dimensional path planning problem
• Decompose the problem into two sub-problems
• Path finding in the workspace
• Refine path to satisfy motion constraints
• When possible, make use of domain specific
  knowledge (such as biomechanics literature in
  this case) or/and use data-driven methods

8
Related work in robotics literature

• Researchers have been trying to solve the problem
  of path planning for highly articulated robots
• The decomposition-based planning approach was
  first suggested by O. Brock and L. Kavraki, ICRA
  2001
• Compute a tunnel in the workspace to capture
  connectivity of free space
• Navigate inside this tunnel using a
  potential-field based approach

9
Related work in robotics literature

• This approach has been demonstrated for a
  free-flying 11 DOF robot
• Motion quality is not a concern as long as an
  obstacle free path to the goal is guaranteed
• Difficult to use in the case of human avatars
  since characterization of the forces acting on
  the avatar is not possible

10
Focus on human-like characters

• Typical human avatar has 60-120 DOFs
• Kinodynamic constraints on human motion in terms
  of velocity, acceleration, turning radius etc.
• Constraints on by how much a joint can rotate in
  its local frame of reference
• Most human motions are severely under-constrained
• Despite extensive research, we are still far from
  producing an accurate biomechanical model of
  human motion

11
Typical character animation framework
(Source J. Kuffner, CMU)
12
Research literature

• All previous approaches to tackling this problem
  can be primarily classified in three categories
• Physically based modeling methods
• Purely data driven methods (using motion capture
  data)
• Use a combination of data-driven methods and
  inverse kinematics to satisfy constraints

13
Model based methods

• Hand designed controllers for human motion,
  primarily locomotion
• Based on the notion of comfort, balance and
  stability to generate plausible motion
• The grasp-based planner, for instance
• It takes a couple of years for a human to learn
  to balance and many more years of experience to
  make intelligent decisions while navigating
• Controller design is daunting and does not work
  well in practice

14
Data-driven techniques

• Motion capture data is very much like video (but
  in 3D)
• Each frame encodes the root position and its
  orientation, and the joint angles in their
  respective local frame of references

15
Planning Biped Locomotion using Motion Capture
Data and PRMs

• Build a PRM (Probability Roadmap) based on
  sampled foot-plant configurations
• Given initial and target constraint, the PRM is
  searched to find a path that is able to connect
  with motion clips
• Motions are adjusted to meet the constraints

Source Planning Biped Locomotion using
Motion Capture Data and PRMs M. Choi, J. Lee,
S.Y. Shin, SIGGRAPH 2003
16
Pipeline

• Sample valid foot-plants randomly in the
  workspace
• Align foot-plants to match start and end
  positions with a pre-recorded motion clip
• Search for a path in an augmented roadmap by
  minimizing a cost function
• Iteratively refine the path (straightening and
  smoothing)
• Body trajectory estimation using Motion
  Retargeting

17
Limitations

• Approach tightly coupled with the PRM constructed
• Dynamic obstacles (such as other moving
  characters) not considered
• Motion retargeting may prove to be expensive if a
  large number of characters are to be animated

18
Fast and accurate goal-directed motion synthesis
for crowds

• Two level synthesis
• Coarse search for global path planning using PRMs
• Finer search for detailed motion synthesis
• Incorporate continuous motion adjustment
• Discrete search to roughly satisfy constraints
• Additional displacements for precision

19
Approach
20
Limitations

• Characters may deviate or wander somewhat from
  the intended goal
• Cannot guarantee collision-free paths in a dense,
  dynamic environment
• Expensive search limits it to offline use
• Running cost directly linked to the number of
  characters being animated and the size and
  complexity of the environment

21
Behavior Planning for Character Animation

• Motions abstracted as high-level behaviors and
  organized into a finite state machine (in
  contrast to motion graphs)
• Build search tree of behavior states and perform
  global planning

Source Behavior Planning for Character
Animation M. Lau, J. Kuffner, SCA 05
22
Behavior Planning for Character Animation
Solution Path(Sequence of Behaviors)
Environment

• Behavior Planner

FSM
Animation
23
Motivation for Pre-computing Search Trees

• Complexity increases with the number of
  characters being animated and the complexity of
  the environment
• Pre-compute possible paths for runtime efficiency
  and handling complex, dynamic environments

Expand all possible states of the FSM
24
Precomputed Search Trees
Runtime
Pre-compute
1) Map Obstacles
Environment
1) Search Tree
Prune and solve for goals by sub-division
FSM
2) Grid-maps
2) Path Finding
25
Discussion

• Interactive synthesis achieved at the cost of
  memory usage
• Planning is no longer A based, use a fast 2D
  bitmap planner
• Re-use same tree in different environments
• Does not guarantee globally optimal solutions
• Size of the tree could be an issue with a varied
  set of motions
• No notion of rules to govern the motion of
  character(s)

26
Motion Graphs

• Dense graphs (unlike FSMs seen earlier) where
  each node is a pose (instead of a motion) and it
  is possible to transition between clips at
  intermediate frames as well
• Transition from one node to the next based on a
  cost function heuristic to generate motion

(Source L. Kovar)
27
Using Motion Graphs for navigation

• Overlay the motion graph onto the environment
  (4D), compute the SCC in the graph and compute
  valid path
• Tightly coupled with the environment, cannot
  handle dynamic obstacles and rapidly becomes
  intractable

Source Evaluating Motion Graphs for Character
Navigation P.Reitsma, N.Pollard. SCA 2004
28
Synthesizing Animations of Human Manipulation
Tasks

• Use a RRT planner to perform path planning for
  the object being manipulated (in 6D C-space)
• Use inverse kinematics to position character and
  simultaneously satisfy all external constraints
• Bias the IK procedure towards poses generated
  from pre-recorded motion capture data
• Post process by path smoothing and computing
  velocity profiles

29
Discussion

• Combines ideas from both model based methods and
  data-driven methods
• Variations in character kinematics handled
• Variations in the kind of manipulation task being
  performed is handled
• Decoupling object position from character pose,
  may not yield a feasible solution (is no
  representative poses are available for
  interpolation)
• Offline computation, cannot be used in
  interactive applications for large number of
  characters

30
Computer Games and other apps

• Widely used in many interactive applications,
  such as SecondLife and computer games
• Games perform navigation for NPCs (Non-player
  characters)
• Discretize the environment into grids and use a
  path finding algorithm (A and its variants) to
  find a path to the goal
• Use a library of carefully selected, either
  keyframed or motion captured moves for playback
  and solve for external constraints such as
  footholds on rough terrain etc. using IK

31
Take away lessons

• The problem is not trivial and there is no
  right way to solve this
• Different applications have different needs and
  require different degrees of correctness in
  synthesizing motion
• Do not think of motion capture data as a panacea
  but use this in conjunction with other methods
  for motion synthesis

32
References

• Precomputed Search Trees Planning for
  Interactive Goal Driven Animation M.Lau,
  J.Kuffner
• Construction and optimal search of interpolated
  motion graphs A.Safonova, J.K.Hodgins
• Planning biped locomotion using motion capture
  data and probabilistic roadmaps M.Choi, J.Lee,
  S.Y.Shin
• Fast and accurate goal-directed motion synthesis
  for crowds M.Sung, L.Kovar, M.Gleicher
• Behavior Planning for Character Animation M.Lau,
  J.Kuffner
• Synthesizing Animations of Human Manipulation
  Tasks K.Yamane, J.Kuffner, J.K.Hodgins
• Decomposition-based Motion Planning A Framework
  for Real-time Motion Planning in High-dimensional
  Configuration Spaces O.Brock, L.E.Kavraki

33
Appendix
Foot-plant transformation
34
Appendix
Augmented Roadmap
35
Appendix
Body Trajectory Estimation using Motion
Retargeting
36
Appendix
Coarse-Level Planner
Repeatedly select sub-goal and run each sub-case