[mlpack] Robotic Arm - GSoC Project Idea

[Marcus Edel]

Hello Vaibhav,

thanks for the update.

> # First we need to find the required final coordinates of the object (in the
>   simulation we would already know that, but in the real world, we need to use
>   some 3D camera).
> # Then we can find the optimal path using trajectory planning techniques
> # After that we can find the joint angles using inverse kinematics
> # Using these joint angles, we can train our model

That is definitely one way, but perhaps we can simplify the way, e.g. by using
only a single camera, this makes the overall training more challenging since,
the preprocessing step is more complex; but at the end, you can directly use the
pipeline for a broad range of applications, without extracting a lot of model-
specific information. Let me know what you think.

> I wanted to you ask how to train mlpack models on large datasets on cloud. It
> would help me to create a detailed timeline for my proposal. Please let me know
> what you think.

Do you mean distributed training? I think for this project we can start small,
but we could we can use Downpour SGD to distribute the load.

Thanks,
Marcus

> On 12. Mar 2018, at 09:18, Vaibhav Jain <vabsweb at gmail.com> wrote:
> 
> Hey Marcus,
> I think this would be a good approach to work on this project:
> # First we need to find the required final coordinates of the object (in the simulation we would already know that, but in the real world, we need to use some 3D camera).
> # Then we can find the optimal path using trajectory planning techniques
> # After that we can find the joint angles using inverse kinematics
> # Using these joint angles, we can train our model
> 
> For learning manipulator kinematics/dynamics, lectures by Prof. Oussama Khatib at Stanford, are very good. The book "Introduction to Robotics: Mechanics and Control" by J. Craig is also very good.
> For simulation, the environments you suggested would be surely helpful. Although, I need to work a little bit more to see the extent to which they can be used.
> 
> I wanted to you ask how to train mlpack models on large datasets on cloud. It would help me to create a detailed timeline for my proposal.
> Please let me know what you think.
> 
> On Wed, Feb 28, 2018 at 6:22 PM, Marcus Edel <marcus.edel at fu-berlin.de <mailto:marcus.edel at fu-berlin.de>> wrote:
> Hello Vaibhav,
> 
> OpenAI released a couple of robotics environments wich could be interesting:
> https://github.com/openai/gym/tree/master/gym/envs/robotics <https://github.com/openai/gym/tree/master/gym/envs/robotics>
> 
> Best,
> Marcus
> 
>> On 27. Feb 2018, at 17:00, Vaibhav Jain <vabsweb at gmail.com <mailto:vabsweb at gmail.com>> wrote:
>> 
>> Hello Marcus,
>> Thanks for this resource. I will sure check it out. In the meantime, I will also research some other previous similar projects.
>> 
>> 
>> On Feb 27, 2018 03:36, "Marcus Edel" <marcus.edel at fu-berlin.de <mailto:marcus.edel at fu-berlin.de>> wrote:
>> Hello Vaibhav,
>> 
>> thanks for the input, I agree ROS is definitely a good option to get started,
>> the available resources is just one point. My initial idea was to integrate the
>> simulator in the OpenAI Gym framework, mlpack already follows the gym interface
>> so the interaction is already in place, but I think we could use the Gazebo
>> extension for Gym to get it working. This looks promising
>> (https://github.com/erlerobot/gym-gazebo <https://github.com/erlerobot/gym-gazebo>). Not sure the install process is
>> straightforward, in the best case we can install everything with a few clicks
>> and most packages are available for a bunch of distributions.
>> 
>> Thanks,
>> Marcus
>> 
>> __
>> Vaibhav Jain
>> 
> 
> 
> 
> Regards, 
> -- 
> - Vaibhav Jain

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://knife.lugatgt.org/pipermail/mlpack/attachments/20180312/9cdb62a8/attachment-0001.html>