Hi, are there still plans for the free-release of MuJoCo Sim for Linux? At the moment, I'm (struggling a lot) trying/figuring out how to reproduce the results from your papers using Openbody/sim. Kind regards, Aj
Yes, MuJoCo Sim will be a free player for Windows, Linux and OSX. It is one of several things I am trying to finish (the main thing missing is the documentation). However Sim is just a player, allowing you to load models and shake them around and explore their dynamics. You will need the Pro version to interface with it programmatically and do things similar to our research. Alternatively, you could try HAPTIX which is free. It has a socket API which is slow and limited compared to the shared-memory model in Pro, but depending on what you are trying to do it may be sufficient. Which results are you trying to reproduce?
Hi Emo, I'm interested in nearly all of your papers/algorithms! At the moment the concrete problem I'm working on is, reproducing/playing with the approach in the paper, Embed 2 Control : A Locally Linear Latent Dynamics Model for Control from Raw Images http://arxiv.org/abs/1506.07365, by Martin Riedmiller's students video here, This papers something of a moonshot - basically going from "Pixels to Torques"! It's fairly general though, and uses 3 neural networks, encoder-decoder-transition, for system identification/linear approximation and control. The way it's setup, and the way I've coded it allows for plug and play/experimentation with different stochastic control algorithms. I've got the silly/simply cart-pole example working, using iLQR, coded in Julia at the moment, and my implementation of the DRAW encoder/decoder, in Torch . There are a few ways I'd like to try and extend this, First, (the most fun bit, it would be nice to try and extend this by using a realistic multi-joint model, with contacts - basically to see if it's any good for real problems, or just a toy/cute idea? I'm really interested in human locomation, (in particular running rather than walking). It's turning out to be rather difficult & involved to do this with all the two physics simulators I've tried so far, (Blender game engine/Bullet/MakeHuman, and more realistically Opensim/Simbody) - I've hit a brick wall basically - no progress? It's frustrating as I don't really need much for the E2C setup, just input the control vector, and costs/rewards, and then render two images. Second, closely related to the the first, a more involved extension I'm working on is to throw away the trained neural network decoder, and use a hard coded physics simulator (i.e. MuJoCo or OpenSim) as the decoder. This builds on the approach outlined in the paper Picture: A Probabilistic Programming Language for Scene Perception, where they use Blender traces programs as their decoder. I hav'nt got this working yet - I tried to do it with Blender and it's game engine, (run by Bullet Physics), but it turned out to be rather confusing. I think it's a good idea though, and I guess it would have a bigger performance improvement than using a more sophisticated control algorithm? Third, I'd like to try and experiment with other control algorithms, the E2C paper uses AICO. I'd like to try Sergey Levines, IOC, Inverse Optimal Control for Humanoid Locomotion. At the moment I think this is the right order to extend my work/experiments. It would really fun & interesting to try and get MuJoCo to work on these projects. I guess a lot of the problems & questions I would have & ask would be useful for you when writing your documentation, and pre-release testing - I'd be happy to write a "dummies/how to get started" guide? Looking at all the simulators available in this review, clearly MuJoCo or OpenSim, is best for what I'm trying to do at present, and what I'm interested in. Basically, at the moment getting a simulator like MuJoCo or/and OpenSim working, and rendering images is the main thing holding me back Be really interested in what you think of these projects? Best regards, Aj
Sounds like you are having lots of fun To answer the earlier question, you definitely need the Pro version for this kind of work. It combines rendering and simulation in one library, making it very convenient for pixel-to-torque learning. I will let you know when the preview version is ready.