Priama syntéza pohybu s použitím sekvenčného Monte Carlo Prezentácia vedeckej práce Rastislav Hekel, Martin Slavkovský

Online Motion Synthesis Using Sequential Monte Carlo Perttu Hamalainen, Sebastian Eriksson, Esa Tanskanen, Ville Kyrki, Jaakko Lehtinen

Introduction Model-Predictive Control (MPC) system for online synthesis of interactive and physically valid character motion 3D human character model can balance, dodge projectiles and return to its given pose The character can improvise a get up strategy after it has been forced to fall

Example of the Synthesized Motion

Overview System generates trajectories of character control parameters for the near future using Sequential Monte Carlo Sampling Samples are generated by multimodal, tree- based sampler Each sample is evaluated by fitness function Best sample is the control strategy for the current frame Maintaining multiple strategies is crucial for adaptation to dynamically changing environments

Goals Time-varying control strategy that drives the character towards the specified goals, while accounting for changes in the environment Motion should be creative and natural with minimal input data System should operate at an interactive frame rate at design time, enabling rapid iteration of the goals and constraints

Comparison to Related Works Work focuses on optimization based animation of active characters instead of passive ragdolls Problem is solved online – offline systems exist for more complex motions System does not require handcrafted state machine or dataset of reference motions Longer planning horizon (4s), multimodal fitness function, more complex character

Contributions The introduction of SMC to online synthesis of physically valid character motion A novel sequential sampling method that allows easy integration of machine learning - the sampler utilizes kD-trees for adaptive sampling Online, near-real-time synthesis of complex get up strategies

Sequential Monte Carlo Sampling Widely used in tracking problems Tracked probability density is approximated by set of samples Samples are weighted and resampled Heavier samples produce more offspring New samples are drawn from proposal densities based on the previous samples

Algorithm Overview

2D Example

Adaptive Importance Sampling Using a kD-tree

Naive vs Gaussian kD-tree

Testing Results The system shows considerable creativity in adapting to surprising situations and utilizing the environment The main drawbacks of the system are that movement is sometimes stiff and has unnecessary joint contortions Method was tested by 1.Throwing spheres at the character 2.Adding sudden impulses to body parts to disturb balance and throw the character around 3.Triggering simulated explosions that add impulses to all body parts

Conclusions Paper has demonstrated that Sequential Monte Carlo (SMC) sampling is a viable approach for online synthesis of complex human motion Sequential kD-tree sampler has suprisingly high performance Sampler is simple enough to be implemented from scratch Improving performance and controlling the style of synthesized movement are the two main items for future work Autors plan to investigate whether sequential sampling is competitive also in offline synthesis, where the function landscape changes over time when the animator interactively adjusts parameters

Thank you for listening