A Differentiable Physics Engine for Deep Learning in Robotics
Authors: Jonas Degrave, Michiel Hermans, Joni Dambre, Francis wyffels
ICLR 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this paper, we test whether these engines are also differentiable and whether this gradient is computationally tractable. We will show how this method does speed up the optimization process tremendously, and give some examples where we optimize deep learned neural network controllers with millions of parameters.Our results show the first prototype of a differentiable physics engine based on similar algorithms as those that are commonly used in modern robotics simulators. |
| Researcher Affiliation | Academia | Jonas Degrave, Michiel Hermans , Joni Dambre & Francis wyffels Department of Electronics and Information Systems (ELIS) Ghent University i Minds, IDLab Technologiepark-Zwijnaarde 15, B-9052 Ghent, Belgium {Jonas.Degrave,Joni.Dambre,Francis.wyffels}@UGent.be |
| Pseudocode | No | The paper describes algorithms and implementation details but does not provide structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide an explicit statement or link for the source code of the described methodology. The provided links are for a video and a Twitter post, not source code. |
| Open Datasets | No | The paper uses custom-built models and simulations (e.g., 'model of a giant soccer ball', 'quadrupedal robot model', 'robot arm model', 'pendulum-cart system') for its experiments, and does not refer to or provide access information for any publicly available datasets. |
| Dataset Splits | No | The paper does not explicitly provide training, validation, or test dataset splits. Experiments are conducted in simulations, optimizing parameters over a simulated time duration. |
| Hardware Specification | Yes | We evaluated both on CPU (i7 5930K) and GPU (GTX 1080) |
| Software Dependencies | No | The paper mentions implementing the physics engine in Theano and references other libraries like Caffe and TensorFlow, but does not specify version numbers for any of these software dependencies. |
| Experiment Setup | Yes | We choose an objective which is evaluated at every time step and then averaged, rather than at specific points of the simulation. The value of the gradient is decreased by a factor α < 1 at every time step. We initialize the controller intelligently. We constraint the size of the gradient to an L2-norm of 1. As a controller, we use a dense neural network with 1 input, 2 hidden layers of 128 units with a rectifier activation function, and 4 outputs with an identity activation function. We use gradient descent with a batch size of 1 robot for optimization, as the problem is not stochastic in nature. The parameters are optimized with Adam s rule (Kingma and Ba, 2014) with a learning rate of 0.001. |