FTL-project

Note: Refer to the results branch of this repository to get the detailed analyis of the experimental results obtained by us.

Simulating hair and fur is very common in the field of Computer Graphics. One of the main challenges for the simulations of these objects is to guarantee a zero extension in their length, since they are supposed to be non-deformable (Think of how things would be if you pull your hair along their length and they keep elongating like a spring!). One method to solve is problem is an extension of a technique called Follow the Leader (FTL) (http://robotics.stanford.edu/~latombe/papers/knotmaking/paper.pdf)

A method for solving this problem of Zero-stretch for dynamic simulations based on the FTL method has been presented as Dynamic Follow the Leader (DFTL) (https://matthias-research.github.io/pages/publications/FTLHairFur.pdf). The main issues with the method given in that paper are as follows:

• No proof of correctness was given for this method
• The velocity correction term actually gives an improper steady state velocity
• An improper Energy curve
• A lot of Artificial damping

This work deals with tackling with these problems to improve the DFTL give an algorithm which solves the zero extension problem, is physically accurate (in the limits of the step size time tending to zero) and does not have such high artificial damping. The results of the experimentations done by us can be found in FTL_Analysis_results.pdf. Moreover, we also give the proof of correctness for the DFTL method; proving that the method would correspond to physically accurate solutions, for particular parameter settings, with the time-step tending to zero. The proof can be found in DFTL_anaylsis.pdf

About the codes

All the codes are built on the same situation of an inextensible thread, assumed as a collection of some n number of particles, the distance between each of them to be fixed as d. The mass of each particles is m and the Gravity is taken as g. The user inputs the total number of particles, the time for the simulation, and the step size.

You can find the codes related to the FTL implementation in the directory FTL_Codes and some other standard time-integrator implementations of the given problem in the directory Other_Codes

About the Videos

All the outputs are given on the situation of a thread of length 10m, consisting of 11 particles, with the mass of each particle being 1 Kg. The gravity constant is taken as 9.8 m/s2. The Spring constant is 1000 SI units and the damping constant is 5 SI units. Also, the time step is 0.05s and the total time for the algorithm to run is 8 or 10 seconds. The names of the videos correspond to their respective algorithms. Also, the Default frame rate corresponds to 100 images per frame. But for specifying different values other than the default ones, the following conventions are used:

• TS/TimeStep: The time step for the given simulation result is the value specified after this (In s)
• m: The mass of the particles for the given simulation is the float value given after this (In Kg)
• r: The distance between consequtive particles on the string is the value described after it (In m)
• S: The damping parameter s_damping for the DFTL simulation is the value given after this
• N: The total number of particles in the string for the simulation (Including the hinged particle)
• TotalTime: The total duration of the simulation is speicified after this (In secs)
• ks: Spring force constant between each consequtive particles as per the model assumed
• kd: Damping force constant corresponding to each springs, as per the considered model

Summary of all the new algorithms for FTL

DFTL_Quad

• In the update rule of v_i = (p_i - x_i)/(Δt) + s_damping(-d_i+1/Δt), the s_damping is varying as f(t) (such that it's value is closer to one when t→0 and closer to zero when t→T) rather than being a constant as in DFTL_Orig
• In some cases, achieves a proper steady state velocity because of s_damping tending to 0 in the last time frames
• Gets a nearly monotonus energy curve for the above situations, as observed experimentally
• The method looks natural because of continuous varations in the s_damping
• An issue that the value of s_damping is dependant on t and not on the x,v which makes it less flexible to external impulses or sudden external motions or interrupts given by the user
• Another issue is that there is no theoretical justification for this. And it is more like a Band-aid method.

FTL_Mem

• Rather than the velocity correction equation v_i = (p_i - x_i)/(Δt) + s_damping(-d_i+1/Δt) as used in the original implementation of DFTL, the correction is added to the position update, which becomes: p_i = x_i + v_iΔt + f_i(Δt)² - s_damping(d_i+1 + d_i)/2
• Has this benefit of theoretical soundness because of doing the corrections prior to the FTL-update as compared to the DFTL implementation which does this correction posterior to the FTL-update
• Achieves a proper steady state velocity, because of the prior corrections
• Monotonously decreases energy for s_damping=1. Does not have for lower values of it, since the method then tends to be behaving like the typical FTL
• More stable and damped than DFTL
• The method looks natural and behaves much like the DFTL method, but achieves a good theoretical improvement over it

FTL_Combine

• A DFTL D₁ (with s_damping as s₁) is run and the total energy of the system E_total is calculated and then it is compared with the initial Energy E₀ and then we check for |E_total-E₀| < C: if it's true then we continue with the next time-step, else we run the DFTL D₂ (with s_damping as s₂) which will be unstable in nature and will help to increase the energy
• Does NOT achieve a proper steady state velocity
• Does NOT have a proper monotonous Energy curve
• Has lesser damping than DFTL, but it may get unstable in some circumstances because of the unpredictable nature of the DFTL plots
• The method seems unnatural because of the sudden bursts of Energies involved

DFTL_Quad_Combine

• Combining the principles of DFTL_Mem and DFTL_Quad and having both the position and velocity corrections as well as their corresponding coefficients (s_p and s_v respectvely) being time-varying as per the corresponding functions f_p(t) and f_v(t) respectively
• Achieves a proper steady state velocity
• Does NOT follow a monotonous Energy curve and shows a peak in the Energy curve in th beginning
• More stable and Damped than DFTL (In the physically relevant implementations)
• The method looks Natural and behaves much like the DFTL method

DFTL_Blend

• Blending two different DFTLs with s_damping=0 and s_damping=1 respectively, as per applying binary search on a parameter α as described in https://www.cs.utah.edu/~ladislav/dinev18stabilizing/dinev18stabilizing.html
• But it is NOT possible to blend two different DFTLs since both of them will be satisfying the string length constraints, and so for their blend to also satisfy this constraint; we would require that α is either 0 or 1