Deep Learning Quadcopter Control via Risk-Aware Active Learning

Authors: Olov Andersson, Mariusz Wzorek, Patrick Doherty

AAAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the efficacy of the approach on a difficult collision avoidance problem with non-cooperative moving obstacles. Our findings indicate that the resulting neural network approximations are least 50 times faster than the trajectory optimizer while still satisfying the safety requirements. We demonstrate the potential of the approach by implementing a synthesized deep neural network policy on the nano-quadcopter microcontroller.
Researcher Affiliation Academia Olov Andersson, Mariusz Wzorek, Patrick Doherty {olov.a.andersson, mariusz.wzorek, patrick.doherty}@liu.se Department of Computer and Information Science Link oping University, SE-58183 Link oping, Sweden
Pseudocode No The paper includes 'Figure 1: Training procedure for deep neural network policy approximations,' which is a flowchart, but no structured pseudocode or algorithm blocks are present.
Open Source Code No The paper mentions using Tensorflow (www.tensorflow.org) and the open-source Crazyflie platform (www.bitcraze.io), but does not explicitly state that the authors' own implementation code for the methodology is open-source or provide a link to it.
Open Datasets No The paper states, 'Training batches contained about 500 000 examples from the trajectory optimizer, which for our problems took approximately 12 hours to generate,' indicating the data was generated by the authors using a trajectory optimizer rather than being from a publicly available dataset with concrete access information.
Dataset Splits No The paper mentions 'early stopping' during training, which implies the use of a validation set, but it does not provide specific dataset split information (e.g., percentages or sample counts for training, validation, and test sets).
Hardware Specification Yes The networks were implemented in Tensorflow1, a graph-based language for numerical computation, and trained using a consumer Geforce GTX970 GPU.
Software Dependencies No The paper mentions software like 'Tensorflow' and 'Free RTOS' but does not provide specific version numbers for any key software components or libraries.
Experiment Setup Yes We use three hidden layers for each problem size, with the network architectures 10-200-200-200-2 and 18-400-400-400-2 respectively. ... Training batches contained about 500 000 examples ... We used the ADAM (Kingma and Ba 2015) stochastic gradient algorithm with mini-batches of size 500 and early stopping. Even just a small amount of dropout (Srivastava et al. 2014), e.g. 1-5%, seemed to be helpful for the larger problems.