From Virtual Demonstration to Real-World Manipulation Using LSTM and MDN
Authors: Rouhollah Rahmatizadeh, Pooya Abolghasemi, Aman Behal, Ladislau Bölöni
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experimental studies validate the three contributions of the paper |
| Researcher Affiliation | Academia | Department of Computer Science University of Central Florida, United States rrahmati,pooya.abolghasemi,lboloni@cs.ucf.edu, abehal@ucf.edu |
| Pseudocode | No | The paper describes the approach in text and with diagrams (Figure 1, 2, 3, 5), but does not include any pseudocode or algorithm blocks. |
| Open Source Code | No | The paper includes a link to a video demonstration (https://youtu.be/9vYlIG2ozaM) but does not provide concrete access to the source code for the described methodology. |
| Open Datasets | No | The paper states, 'The demonstrations were collected from a single user, in the course of multiple sessions. In each session, the user performed a series of demonstrations for each task.' This describes a custom-collected dataset, but no access information (link, DOI, citation) is provided for public availability. |
| Dataset Splits | Yes | We use 80% of the data for training and keep the remaining 20% for validation. |
| Hardware Specification | No | The paper mentions using 'a Rethink Robotics Baxter' for physical experiments but does not provide specific details on the hardware (e.g., GPU models, CPU specifications) used for training the neural network. |
| Software Dependencies | No | The paper mentions 'Unity3D game engine' for simulation and 'Microsoft Kinect sensor' for vision, but does not provide specific version numbers for software dependencies used in the neural network training or implementation. |
| Experiment Setup | Yes | The network is unrolled for 50 time steps. All the parameters are initialized uniformly between -0.08 to 0.08 following the recommendation by (Sutskever, Vinyals, and Le 2014). Stochastic gradient descent with mini-batches of size 10 is used to train the network. RMSProp with initial learning rate of 0.001 and decay of 0.99-0.999 (based on number of examples) is used to divide the gradients by a running average of their recent magnitude. In order to overcome the exploding gradients problem, the gradients are clipped in the range [-1, 1]. We stop the training when the validation error does not change for 20 epochs. |