The last two weeks was more of experimenting with various approaches of controling F1-car and Iris drone in DL-Studio and Behavior Metrics.
- Stacked PilotNet experiments
- DeepPilot CNN with drones
- Concluding my GSoC journey and starting my member collaboration!
Stacked PilotNet experiments
The stacked PilotNet setup is interesting because it is aimed towards learning temporal relationships without using a memory based network setup like LSTMs. There are a lot of ways to incorporate the set of frames which I would like to explore:
Case 1: Spanning wide across the horizon
In the previous experiments as demonstrated in one of my previous blogs  the consecutive set of last H-frames are stacked for a H-step stacking and the PilotNet module acts on the setup. While this works but given the frequency of receiving frames (20 Hz), the last H-steps are somewhat very same. So, this makes so siginificant difference from working with a single frame.
This case aims towards having a wider span of the horizon, which means that instead of taking H-consecutive frames, we can choose H-last frames at an interval of M frames. Now, increasing this M to a lot fails the experiments, however a value of 5 tends to work. After analyzing, it seemed that the even M=5 does not significantly changes the frames.
Case 2: Considering deviations motivated from explicit brain
The second case is more motivated by the explicit PD brain which works on the error in deviation of the red-line from the middle of the frame. This was we can decide the frames to be stacked based on difference in the deviations.
DeepPilot CNN with drones
Finally, with the drone dataset   and the DeepPilot implementation in Deep Learning Studio , I was able to implement a CNN based controller for the Iris drone. The CNN is trained on the dataset and the controller is implemented as a modified DeepPilot with only outputs as the forward velocity, rotation velocity and vertcal velocity. However, the controller is not yet working and fails to learn the correlation of the z-velocity with the red-line width in the image frame. This is a future work. A pull request was made to the Behavior Metrics repository which allows us to test the trained deeppilot networks.
Concluding my GSoC2021 journey
Finally, my journey as GSoC2021 student with JdeRobot ends with a lot of experiments and learning. I am very happy with the progress I made and I am looking forward to working with the DL-Studio team as a member of the JdeRobot community. I am thankful to the JdeRobot community for giving me such a chance to work on such an interesting project. I am also fortunate that they believed in me and allowed me to work directly on the main repository. Now, with this work:
- Behavior Metrics has both PyTorch and TensorFlow implementations of the brains.
- DL-Studio has a new implementation of the PilotNet module which is based on the base PilotNet paper.
- There is an added PilotNet stacked setup ready for use in DL-Studio.
- Finally, I was able to introdice Iris drone into Behavior Metrics and DL-Studio with JdeRobot Drones and DeepPilot implementations.
You can find all my videos for this GSoC project in the playlist . I am looking forward to a very productive collaboration as a member of such a knowledgeable community.
Issues and Pull Requests
Pull Requests Created
Pull Requests Closed