Perceptual Locomotion with Controllable Pace and Natural Gait Transitions Over Uneven Terrain

Codebase for paper Perceptual Locomotion with Controllable Pace and Natural Gait Transitions Over Uneven Terrain.

Setup

Our repository works for Python 3.8+. Install dependencies and source in developer mode:

python -m pip install -e .

Training heightmap eutoencoder

First, generate the data used in training the heightmap autoencoder:

python blind_walking/examples/hover_robot.py -n 1800

A few notes on the script:

• -n 1800 sets the number of timesteps collected to be 1800.
• --record to record the hovering robot's trajectory over different terrains.

Shift the generated heightmap.npy data file to a folder ./blind_walking/examples/data/heightmap.npy. Then, train the heightmap autoencoder:

python train_autoencoder.py

Loading an Existing Model

A pre-trained locomotion policy is provided in saved_models/policy.
To run the policy in Pybullet simulation in headless mode:

python scripts/enjoy_with_logging.py --algo ppo --env A1GymEnv-v0 -f saved_models/policy --no-render

The --record flag can optionally be added to record the episode as a video. Video parameters such as camera angles can be configured in blind_walking.envs.locomotion_gym_config.SimulationParameters.

Training a New Model

To reproduce the training settings from the paper:

python train.py --algo ppo --env A1GymEnv-v0 -f logs --n-timesteps 4000000

Acknowledgements

This codebase draws inspiration from the following codebases:

• Stable Baselines3 Zoo
• Motion Imitation