learned_cost_map

This repo contains code to train and deploy a learned traversability costmap.

Requirements

This code was tested with the following package versions:

• Python 3.8
• PyTorch 1.10 (CUDA 11.3)
• NumPy 1.22

Deployment

In order to run the ROS node that generates the learned costmap, we first need to make sure that we have three topics being published:

• a top-down heightmap (in our case, it is /local_height_map_inflate)
• a top-down rgbmap (in our case, it is /local_rgb_map_inflate)
• an odometry (in our case, it is /integrated_to_init)

The top down maps should follow the dimensions set in learned_cost_map/configs/*map_params.yaml. If the top-down maps are obtained using TartanVO, then we need to make sure TartanVO is running. To run TartanVO, run the multisense_register_localmapping.launch inside the physics_atv_deep_stereo_vo package.

Once the necessary topics are published (or TartanVO is running), we can run the learned_cost_map.launch launch file in this package, which will publish an occupancy grid with the costmap.

The current dimensions of the costmap are 12m x 12m with a resolution of 0.02 m and the origin at (-2, -6).

The trained models are saved inside learned_cost_map/models. The output costmap will be published to '/learned_costmap'

Training

Processing Rosbags

TODO

Labeling with pseudo ground-truth traversability costs

TODO

Labeling with IMU cost function

TODO

Collecting statistics of dataset

TODO

Data Balancing

TODO

Training

TODO

Visalization

TODO