Visuomotor policies from event-based cameras through representation learning and reinforcement learning. Accompanies our paper: https://arxiv.org/abs/2103.00806
Contains event data from AirSim, trained eVAE weights and trained obstacle avoidance policies in three folders:
evae_data: Sample data from 'poles' and 'gates' environments to train eVAE representations over.
weights: Sample eVAE representation weights trained over 'poles' environment, one with XY data only (evae_xy) and one with full event data (evae_full_tc).
policies: Policies trained using both kinds of eVAE weights as above: brp_full (full event data) and brp_xy (XY data only)
Place evae_data in the event_vae folder and the rest in event_rl folder.
Windows binaries for AirSim drone obstacle course.