使用Drake进行动作捕捉问题是否是个好主意?

huangapple go评论156阅读模式
英文:

Is it a good idea to use Drake for motion capture problems

问题

I'm searching for a toolbox that will aid in the development of a trajectory optimisation method to perform pose estimation/motion capture of quadruped animals. I have data from multiple cameras and would like to fit a skeletal model to the data. I'm using trajectory optimisation to obtain the pose and dynamics (internal torques and external contact forces) of the animal during locomotion.

I have a solution using custom software -- but it is very slow and hinders the experimentation process.

Would you recommend using Drake for this? If yes, which modules do you recommend I take a look at? If not, what are the drawbacks of trying to use Drake and do you recommend any other software for this?

英文:

I'm searching for a toolbox that will aid in the development of a trajectory optimisation method to perform pose estimation/motion capture of quadruped animals. I have data from multiple cameras and would like to fit a skeletal model to the data. I'm using trajectory optimisation to obtain the pose and dynamics (internal torques and external contact forces) of the animal during locomotion.

I have a solution using custom software -- but it is very slow and hinders the experimentation process.

Would you recommend using Drake for this? If yes, which modules do you recommend I take a look at? If not, what are the drawbacks of trying to use Drake and do you recommend any other software for this?

答案1

得分: 0

为了确保我理解正确,您好像有一个您喜欢的模型(运动学 + 动力学参数已固定),并且您有一些标记点的位置轨迹,您想要恢复扭矩和地面反作用力?您是否考虑使用轨迹优化,其中包括匹配标记点随时间变化的成本?

轨迹优化代码的选择在于您是否知道每一帧上哪些物体发生了接触,或者您正在尝试推断这一点。对于已知接触序列的情况,您可以查看这个示例或者尝试稍微通用一点的混合多体代码,它应该很快就会发布。

如果您正在推断模式序列,那么您可能需要类似于这个链接或通过接触进行迭代LQR的方法。人们确实在Drake中实现了这些方法,但是Drake的主要存储库中尚未提供完善的实现。

有关这些算法的一些更一般的注意事项,请参阅这里。

英文:

To make sure I understand, it sounds like you have a model that you like (kinematics + dynamics parameters are fixed), and you have the position trajectories of, for instance, a handful of markers, and you want to recover the torques and ground reaction forces? And you're thinking of using trajectory optimization with a cost on e.g. matching the positions of the markers over time.

The choice of trajectory optimization codes can change somewhat significantly depending on whether you know which bodies are in contact on each frame or whether you are trying to infer that. For the cases of known contact sequences, you might look at this example or try out the slightly more general hybrid multibody code which should land soon.

If you're inferring the mode sequence then you might want something like https://groups.csail.mit.edu/robotics-center/public_papers/Posa13.pdf or iterative LQR through contact -- people have definitely implemented those in Drake but we don't have a shiny implementation in Drake's main repo yet.

There are some more general notes about these algorithms here.

huangapple
  • 本文由 发表于 2023年5月10日 15:58:06
  • 转载请务必保留本文链接:https://go.coder-hub.com/76216133.html
匿名

发表评论

匿名网友

:?: :razz: :sad: :evil: :!: :smile: :oops: :grin: :eek: :shock: :???: :cool: :lol: :mad: :twisted: :roll: :wink: :idea: :arrow: :neutral: :cry: :mrgreen:

确定