-
generate plots to visualize the inertial data collected from the IMU on the spot (analog, fft, PSD)
- The scripts can be found in scripts/viz_spot_inertial.py
- We are now using the PSD instead of raw analog values of IMU, feet and leg achieved better results, so scripts/train_naturl_representations.py has been modified to use PSD instead of raw analog values.
-
save the k-means clustering results to a file. Use the GUI to get preferences from human, and save the preferences
- Skipped the GUI part, but the preferences are now hardcoded in the scripts/train_naturl_cost.py
- To manually get the preferences, look at the saved 25 patch grids and assign the preferences in scripts/train_naturl_cost.py
-
train the cost function network
- We now have trained cost function network for the following 8 terrains : [asphalt, bush, cement, dark_tile, grass, marble_rock, pebble_pavement, red_brick]
- The trained cost function network with 99% accuracy can be found in the folder : models/acc_0.99979_best
-
write a wrapper script to wrap the encoder and the costfunction network into a single .pt file
- We are saving the wrapped model in scripts/train_naturl_cost.py
- scripts/plot_naturl_cost.py contains scripts that can be used to plot the cost function network, using the wrapped model
-
QOL Improvement- Make the NATURL representation learning script (cost/train_naturl_representations.py) train parallely on multiple GPUs.
- Now can train both representations and cost function network parallely on multiple GPUs.
-
Design the experiments. Come up with a list of training terrains and the scenarios we will need to train and test the NATURL algorithm on the spot.
-
Train the encoder and cost function network for the appropriate terrains and sync with Elvin for the actual experiments on the spot.
-
Start writing the pre-writing form for this paper. Start writing the Abstract and Introduction, Related Work and Experimental Setup sections.
-
Start training on hazard / pepi machines