Skip to content

Latest commit

 

History

History
63 lines (54 loc) · 1.32 KB

LBC.md

File metadata and controls

63 lines (54 loc) · 1.32 KB
Raw

Training LBC

This page provides instructions to train a LBC model.

If you find this to be useful, please also cite:

@inproceedings{chen2019lbc,
  author    = {Dian Chen and Brady Zhou and Vladlen Koltun and Philipp Kr\"ahenb\"uhl},
  title     = {Learning by Cheating},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2019},
}

Note: for each stage, you can use wandb to visualize and monitor the progress.

Setup

Create the following config_lbc.yaml file.

---
num_plan: 5
camera_x: 1.5
camera_z: 2.4
camera_yaws: [0,-30,30]
seg_channels: [4,6,7,10,18]
seg_weight: 0.05
imagenet_pretrained: True
log_wandb: True
noise_collect: False
x_jitter: 3
a_jitter: 15
crop_top: 8
crop_bottom: 8
bev_model_dir: [PATH TO PRIVILEGED MODELS]
rgb_model_dir: [PATH TO PHASE MODELS]
main_data_dir: [PATH TO DATA]

Data collection

Pleaes refer to RAILS.md for instructions on data collection.

Stage 1: privileged BEV model

  • Train the model
python lbc.train_phase0 --save-path=[PATH TO SAVE BEV MODEL]
  • Edit config_lbc.yaml
bev_model_dir: [PATH TO SAVE BEV MODEL]

Stage 2: distilled RGB model

  • Train the model
python lbc.train_phase1
  • Edit config_lbc.yaml
bev_model_dir: [PATH TO RGB MODEL]

lbc