Interactive Control of Diverse Complex Characters with Neural Networks

Authors: Igor Mordatch, Kendall Lowrey, Galen Andrew, Zoran Popovic, Emanuel V. Todorov

NeurIPS 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show the performance of our method on a biped walking task in figure 2 under full method case. To test the contribution of our proposed joint optimization technique, we compared our algorithm to naive neural network training on a static optimal trajectory dataset. ... The results are shown in no joint case. ... To test the contribution of noise injection, we used our full method, but disabled sensory and hidden unit noise... Additionally, we have compared the performance of different policy network architectures on the biped walking task by varying the number of layers and hidden units. The results are shown in table 1.
Researcher Affiliation Academia Igor Mordatch, Kendall Lowrey, Galen Andrew, Zoran Popovic, Emanuel Todorov Department of Computer Science, University of Washington {mordatch,lowrey,galen,zoran,todorov}@cs.washington.edu
Pseudocode Yes Algorithm 1: Distributed Stochastic Optimization
Open Source Code No The paper does not provide a specific repository link, an explicit statement about code release, or mention code availability in supplementary materials for the methodology described.
Open Datasets No The paper describes the internal process of generating the dataset used for training: 'To train a neural network for interactive use, we required a data set that includes dynamically changing task s goal state. ... Our trajectory generation creates a dataset consisting of trials and segments.' However, it does not provide concrete access information (link, DOI, formal citation) for a publicly available or open dataset.
Dataset Splits No The paper mentions '1000 training and test trials' and discusses 'training performance' versus 'test performance' (e.g., in Figure 2 and section 8.1), but it does not explicitly describe a separate validation dataset split or a cross-validation setup.
Hardware Specification Yes Amazon Web Service s EC2 3.8xlarge instances provided the nodes for optimization, while a g2.2xlarge instance provided the GPU.
Software Dependencies No The paper mentions 'We used the Mu Jo Co physics simulator [16] engine for our dynamics calculations.' and 'We used a custom GPU implementation of stochastic gradient descent (SGD) to train the neural network control policy.' However, it does not provide specific version numbers for MuJoCo or any other key software libraries or dependencies used in the experiments.
Experiment Setup Yes All our experiments use 3 hidden layer neural networks with 250 hidden units in each layer (other network sizes are evaluated in section 8.1). The neural network weight matrices are initialized with a spectral radius of just above 1, similar to [15, 5]. ... Because our physical state evolution is a result of optimization (similar to an implicit integrator), it does not suffer from instabilities or divergence as Euler integration would, and allows the use of larger timesteps (we use t of 50ms in all our experiments). ... The values of the algorithmic constants used in all experiments are σε = 10 2 σγ = 10 2 α = 10 λ = 102 η = 10 2.