Professional Documents
Culture Documents
Takeaway
Survey of recent research in the field of
interactive animation from the point of view of a game developer. Assessments of the effectiveness and overall applicability of proposed techniques A pragmatic distillation of subsets of ideas from the large, fully designed systems presented.
Emphasis on techniques and insights applicable now.
They remain kinematic, which from the point of view of physical interactions is equivalent to having infinite inertia
the personality and style of their animations, seamlessly interact with a physically simulated world and take orders from the gamepad of abusive players and confused AIs.
Figure 1: Physical feasibility is one of many challenges faced by otherwise content virtual people.
treated as an afterthought
Balance Quality and results are not mentioned Walk with various gaits Not quantifiable? Handle external This is a first order perturbations such as concern for games. varying terrain and being pushed Robustness as measured Transition between by physical simulation animations results is treated as the Approximate animations highest priority result. from motion capture
Integrator receives forces and torques, forward dynamics equations are solved and the world state accordingly
Mixer linearly blends control parameters rather than poses. Similar controllers are synchronized such that they are in the same state. A Balance post process is applied. (see next slide for details)
System is based on PD controllers acting at joints Each animation has its own individual control system that operates the PD gains and targets with values specified by a simple finite state machine. While this is not treated as a first order concern, the
resulting controllers require very little memory. Compression ratios of 100:1 are typical.
Balance control system is an adaption of the original hopper control system developed by Raibert in the 80s. Main considerations are foot placement upon landing and inertial forces due to acceleration of legs in flight
Raiberts Hopper is Simple and effective. Unfortunately it cannot balance when stationary.
[Video]
Simulated characters in simulated worlds require simulated skills. An extension of SIMBICON Machine learning is used to create new animation controllers based on Existing controllers New challenges such as variations in terrain, new goals, etc Learning framework based on numerical optimization employing continuation methods i.e. Creating an animation controller for pushing a table from a walking motion capture.
Continuation Methods:
[Video]
papers, they are considered valid despite the relatively low sample sizes and relatively poor control procedures of the studies.
Given a source walk cycle and model, offline optimization is used to generate many single step walk cycle variations Each variation is generated to with the optimization goal of a random foot placement offset.
Variations are assembled to create what the authors call a Step-to-Step Dynamics Model or an SSDM .
The step-to-step dynamics model (SSDM). The nonparametric (example-based) model makes predictions using the results of the offline synthesis. The given dimensions for the state and actions spaces are for the 2D bipeds.
At run time, the controller uses the SSDM with its reference walk cycle to generate a continuous walk cycle that satisfies constraints discovered in real time. The SSDM itself stores its data as samples compressed via principle component analysis on the differences between learned steps and the reference walk cycle. Given the strong correlation between variations the authors discover the only 2-4 principle components are required. Sampling consists of applying an N dimension blend on reference samples and then combining the result with the reference walk cycle.
Takeaways The overall framework provides: A potential alternative to traditional foot placement IK. Higher quality results A potential alternative to animator provided walk cycle transitions or variations Saves animator time
Take-Away: PCA Hypothesis Principle Component Analysis (PCA) is a technique used for the lossy compression of vector data sets. A 3 DOF analogy might be DXT1 compression of a single quad of pixels. In this case only the first Principle component is maintained and this 3D vector is the is the difference between the two end point colors. The early PCA vectors resulting from all poses of a character are strongly correlated with the signature style and physical characteristics of that character.
Take-Away: PCA Hypothesis When performing inverse kinematics, using PCA vectors as the degrees of freedom and weighting each in a manner inversely proportional to it s Eigenvalue, a least squares IK solution is likely to represent a pose variation that correlates highly with the preferred poses of that character.
Figure 3: 1st Principle component vector consists of DOF values that produce a characteristic pose.
Take-Away: PCA Hypothesis (cont) There are a few conditions that the dataset used to generate the Principle Component Vectors must meet for this hypothesis to hold: Should be representative of the entire range of motions preferred poses Should be representative in a quantitative, statistical sense Appropriate DOF should form the initial basis Forces or Poses? May require doctoring of the input i.e. weighting the std deviation of underrepresented poses.
Figure 4: Animation set has been over-trained. The 1st Principle Component vector is not ideal.
An initial set of animations is collected Animations are time-warped such that each cycle takes 2 seconds PCA is performed on these animations and the 20 top Eigenmodes are maintained These are run through a physics filter , which iteratively modified them until they are physically valid
(cont)
Components The authors claim that 20 Eigen modes is enough for mostly lossless representation New animations are registered into the motion database and stored as linear combinations of the components The authors state the following applications: Effective motion retargeting Kinematics, dynamics and style Physics improves believability of motions Implicitly enforced by the Components Consistent, realistic motion generation from sparse data. Registration enforces the constraints implied by the sparse data and fills in the blanks with motion from existing motions.
Combines kinematic motion with dynamic reactions that are constrained to minimize interference with muscles essential to the base motion. To accomplish this, the system does the following: Projects motions
PCA, these Components act in the null space of the most significant Components. They correspond to motions requiring little or no muscle action Hence the term unactuated Furthermore the composite effect of these components tend to cancel each other out with respect to side effects at the primary, well actuated DOF.
[video]
their style Primary issues is that for relatively strong perturbations the response is not convincing you would expect perturbation to effect the walk cycle as a whole, unfortunately the problem statement forbids this. Changing the problem statement would invalidate this technique.
Uses machine learning to construct physical controllers that can emulate motion capture data while maintaining balance Controllers operate on
an animator specified biped. The system behaves best when the virtual biped is similar to captured subject.
(cont)
construct a finite state machine that describes potential transitions between animations. Further controllers are generated and trained for each transition All animations can transition to the balance/ recovery states.
(cont)
(cont)
Optimization strategy: Initialize solution to {0} Scatter random points about the current solution Optimize each point using a gradient descent
technique Select best point(s) and repeat Slow, brute force method. Given optimization landscape, faster search methods such as Conjugate Gradients and Quasi Newton fail to converge, so the others use a downhill simplex descent In the second phase of learning (transitions) the optimization parameterizations consist of a linear combination of existing controllers Speeds up the optimization considerably Maintains style of source animations Key takeaway.
(cont)
very few biped/locomotion specific heuristics are feasible Blending controllers trained on animation with a specific style seems yield new controllers that maintain this style. Effectiveness demonstrated for transitions, will blending work with more general extrapolations?
Animating natural
demonstrated to human motion in provide an automatic dynamic environments solution to parts of is difficult for various this problem by the reasons, including robotics community. complex geometric In the field of video and physical game animation, we interactions. require style and personality in addition Simulation combined to competent with physical locomotion. controllers has been
Simulation of Stylized
Describes the systematic
(cont)
synthesis of controllers that can reproduce a range of locomotion styles in interactive simulations. Given a reference motion that describes the desired style, a control system is derived that can reproduce that style in simulation and in new environments.
Figure 1
framework for computing a solution that is optimal in the context of any number of prioritized goals.
In this case the goals are balance and
Simulation of Stylized
A balance strategy is precomputed for the given style using automated analysis of linear time varying approximations. By tailoring the balance strategy in this manner, a controller preserves the style better than a more cautious strategy
(cont)
Simulation of Stylized
(cont)
joint angles to compute the accelerations needed to preserve the given style. Reference motions guide both the style and balance feedback. The style feedback aims to preserve the nuances of the motion The balance feedback seeks to adapt the motion of three balance critical segments The control algorithm computes a final set of forces by maintaining a desired tradeoff between the balance and style feedback.
Simulation of Stylized
Dynamic balance is a hard problem Many dof -> infinite solutions Numerically unstable -> most states invalid Valid states may be balanced, but awkward i.e. if you have a tendency to lean to the left, cranking your head to the right may help you to balance but it is not the most natural way to address the issue.
(cont)
Simulation of Stylized
To mitigate these problems, the authors:
Project the higher detail (60 DOF) biped
(cont)
matrix.
Use a Linear Quadratic Regulator (LQR)
stable
Eliminates the need to explicitly
Simulation of Stylized
The 3 link model s root is the contact
(cont)
foot and it has two 3 DOF actuators acting at the hips One for the torso One for the swing leg The linearalized dynamic equations are precomputed, rolled together with the equations necessary for the LQR, factored and stored for each time step. This amounts to about 20k per second of animation The full character would require about 2 meg per second. Aside: Is 60hz necessary?
[Video]
Simulation of Stylized
The results are promising:
(cont)
Maintains lifelike human motion in dynamic environments, which is difficult to accomplish with kinematics or dynamics alone Transforms a single recorded motion, valid for one environment only, into a general purpose action that can be used in many other settings or even composed with other actions to create versatile characters. Performance reasonable in worst case, plenty of room for optimizations in normal situations.
of a physical simulation 6 DOF springs are attached to the root of the dynamic model to keep it from falling over.
While not mentioned explicitly by the authors, blending in world space helps preserve the momentum characteristics of the source animations, thus making the result more physically plausible (if the source animations is physically plausible)
space is the fact that the Hessian describing the system dynamics in world space coordinates varies less across poses than the corresponding Hessian describing the system dynamics in joint space. This insight is arguably the primary take-away from the paper. A natural corollary is that physical controllers will usually be more robust when operating in world space.
The authors point out this observation in the diagram below.
Impact
(cont)
allow for recovery. Intended to bridge the gap between Ragdolls and layered hit reactions for mild impacts that do not affect the characters momentum or trajectory
Animations with different stepping strategies form the basis for reactions Foot placement of animations is modified counteract the characters momentum. This balance strategy is basis for Asimo s control system The ideal foot placement point is called the Zero Moment Point
(cont)
(cont)
Fits in nicely with existing systems Performance cost is low Not too far from what some people do
currently
Motion
Reaction Animation
Impact
q(t)
Impact
Blend
Simulation
Time
Reaction Animation
(cont)
Assumes many hit reactions (paper cited 500 in their tests) Finding the appropriate reaction becomes a bottleneck
Uses an optimized search tree based on a Support Vector Machine (SVM) Quality of SVM is highly dependent on the DOF used in classification Tried many combinations Optimal set of DOF was 20 in total:
Momentum (linear and angular) at various time steps surrounding the impact Center of mass at same time steps Orientation at time of impact
(cont)
[video]
Defensive Motion
Impact
q(t)
Defensive Motion
Blend
Simulation
Time
Reaction Animation
Inspired by the paper Animating reactive motion using Momentum-based Inverse Kinematics
Uses Momentum based IK to drop animation Uses principles of Dynamic Balance/Zero moment point to synthesize steps from example
Predict Momentum
Adjust for situation based on Balance Combine with dynamic constraints and existing motion
Kinematic Playback
potential. When evaluating ideas, it is important to remember that the whole is not always greater than the sum of its parts.
Figure 3: When shipping a game, elegance and simplicity occasionally slip. When publishing a paper, elegance and simplicity occasionally slip.
Characters that preserve the personality and style of their animations, seamlessly
interact with a physically simulated world and take orders from the gamepad of abusive players and confused AIs.
Just about every successful entertainment medium features compelling characters personality and style are essential. Animators are often better at creating compelling characters than random algorithms/equations. In games we have Animators (hopefully) and many algorithms/equations.
Presenting optimal relevant algorithms that preserve the personality and style of characters/animators is the focus of this lecture.
from
Characters that preserve the personality and style of their animations, seamlessly interact with a physically simulated
take orders from the gamepad of abusive players and confused AIs.
world and
Figure 1: Physical feasibility is one of many challenges faced by otherwise content virtual people.
required at any moment due to open ended nature of player actions is non-trivial Rule based controllers handle this via animation selection guided by heuristics.
i.e. pre-authored decision tree populated with animation clips. Many problems and failure cases Still the basis for many animation systems
the last five years, the amount of research dedicated to this problem has grown exponentially.
The Arsenal
Meta techniques
Generalizing from examples
Interpolate and/or extrapolate from source animations Weighted blend of source (convex interpolation) Transform source to a new basis using direct techniques (i.e. PCA) Use source to train/derive more complex models. Splice to micro motion snippets and transition rapidly
Synthesis
Physical Simulation
Rag dolls Secondary dynamics
techniques, as each has its limitations and strengths. Hence the term Arsenal :
Use, understand, build, augment, combine, patent
Perales CASA 08 The aim of this paper is to model characters that have PERSONALITY, feel EMOTIONS and can MANIFEST EMOTIONAL STATES The motive is to create unique, real, distinguishable individuals depending on their personalities and emotional states