Pigeon Head-Bob

Code for reproducing results seen in my bachelor's thesis "Modeling Head-Bobbing in Pigeon Locomotion using Reinforcement Learning" (PDF file). Refer to the thesis for any details regarding the experiments in the repository.

We used the soft actor critic (SAC) implementation in RLkit by vitchyr and a proximal policy optimization (PPO) implementation built on top of RLkit RlkitExtension for training the pigeon models, both of which are included as sub-repositories.

Abstract

Head-bobbing is a behavior unique to forward locomotion in small birds, mainly pigeons, that consists of a hold phase, where they lock the position of their heads into one position, and a thrust phase, where they move them to a different position. 2 main functionalities of the behavior have been proposed in preliminary research: visual stabilization and induction of motion parallax; however, there is a lack of research that focus on validating their sufficiency by attempting to reproduce it in environments that take physics into account. In our research, we construct a simplified model of pigeons that represent their heads, necks, and bodies and validate the preliminary hypotheses regarding the functionalities of the behavior using reinforcement learning.

Pigeon OpenAI Gym Environments

We constructed 2 OpenAI Gym environments that utilize the simplified pigeon model

• PigeonEnv3Joints (code)

  • Pigeon model is tasked to move its head to predefined target locations
  • Head follows a path that represents head-bobbing behavior
  • Reward functions
    • head_stable_manual_reposition
      • Negative distance between the target locations and the position of the head relative to a threshold value max_offset.
    • head_stable_manual_reposition_strict_angle
      • Stricter version of head_stable_manual_reposition
      • Rewards are only produced when the angle of the head-tilt is within 30 degrees.

• PigeonRetinalEnv (code)

  • Reward functions modeled on the 2 functionalities of the head-bobbing behaviors courtesy of the preliminary hypotheses
    • motion_parallax
      • Sum of velocities of external objects within the retina relative to each other
      • Represents motion parallax induction during the thrust phase
    • retinal_stabilization
      • Negative sum of velocities of external objects within the retina
      • Represents retinal stabilization during the hold phase
    • fifty_fifty
      • Equally-weighted sum of retinal stabilization and motion_parallax

Getting Started

Clone the repository.

git clone https://github.com/johnlime/pigeon_head_bob.git

Set the current directory to the repository's main directory.

cd pigeon_head_bob

Add the current directory to PYTHONPATH

export PYTHONPATH=$PWD

Dependency Installation using Anaconda

The following instructions assume that we are training reinforcement learning models in Linux and conducting testing and visualization of them in MacOS.

Linux

Run the following command for installing dependencies for Linux.

conda env create -f conda_env/rlkit-manual-env-linux64gpu.yml

Activate the Anaconda environment using the following command.

conda activate rlkit-manual

MacOS

Run the following command for installing dependencies for the MacOS.

conda env create -f conda_env/pybox2d-rlkit-manual-env-mac.yml

Activate the Anaconda environment using the following command.

conda activate pybox2d-rlkit-manual

Minimal Dependencies for Using Pigeon Environments

Optionally, you can choose to construct an Anaconda environment solely for using the 2 pigeon environments (without the RLkit training).

conda env create -f conda_env/pigeon_minimal_env.yml
conda activate pigeon-env

Dry Runs of Pigeon Environments

Random policies can be run on the environments via the following command.

python run/pigeon_run.py -env <ENVIRONMENT>

• -env, --environment
  • Name of the environment to run
  • PigeonEnv3Joints
  • PigeonRetinalEnv

Training Reinforcement Learning Controllers

Soft Actor Critic

Run the following command for SAC training.

python run/sac.py -env <ENVIRONMENT> -bs <BODY_SPEED> -rc <REWARD_FUNCTION> -mo <MAX_OFFSET>

• -env, --environment

  • Name of the environment to run
  • PigeonEnv3Joints
  • PigeonRetinalEnv

• -bs, --body_speed

  • Body speed of the pigeon model

• -rc, --reward_code

  • Specify reward function associated with the set environment
    • PigeonEnv3Joints
      • head_stable_manual_reposition
      • head_stable_manual_reposition_strict_angle
    • PigeonRetinalEnv
      • motion_parallax
      • retinal_stabilization
      • fifty_fifty

• -mo, --max_offset

  • Specify max offset for aligning head to target
  • Not necessary for PigeonRetinalEnv
  • Is set to 0.0 by default

• The resulting data are stored under src/rlkit_ppo/data/

Proximal Policy Optimization

Run the following command for PPO training.

python run/ppo.py -env <ENVIRONMENT> -bs <BODY_SPEED> -rc <REWARD_FUNCTION> -mo <MAX_OFFSET>

• -env, --environment

  • Name of the environment to run
  • PigeonEnv3Joints

• -bs, --body_speed

  • Body speed of the pigeon model

• -rc, --reward_code

  • Specify reward function associated with the set environment
    • head_stable_manual_reposition
    • head_stable_manual_reposition_strict_angle

• -mo, --max_offset

  • Specify max offset for aligning head to target

• The resulting data are stored under src/rlkit_ppo/data/

Running the Reinforcement Learning Policies on the Pigeon Environments

Run the following command after training the reinforcement learning policies.

python run/pigeon_run.py -env <ENVIRONMENT> -dir <DIRECTORY_PATH> -bs <BODY_SPEED> -rc <REWARD_FUNCTION> -mo <MAX_OFFSET>

• -env, --environment

  • Name of the environment to run
  • PigeonEnv3Joints
  • PigeonRetinalEnv

• -dir, --snapshot_directory

  • Path to the snapshot directory is within src/rlkit_ppo/data/

• -bs, --body_speed

  • Body speed of the pigeon model

• -rc, --reward_code

  • Specify reward function associated with the set environment
    • PigeonEnv3Joints
      • head_stable_manual_reposition
      • head_stable_manual_reposition_strict_angle
    • PigeonRetinalEnv
      • motion_parallax
      • retinal_stabilization
      • fifty_fifty

• -mo, --max_offset

  • Specify max offset for aligning head to target
  • Not necessary for PigeonRetinalEnv
  • Is set to 0.0 by default

• -v, --video

  • Export result to video

Visualizing the Head Trajectories Generated by the Controllers

Run the following command.

python run/pigeon_headtrack.py -env <ENVIRONMENT> -dir <DIRECTORY_PATH> -bs <BODY_SPEED>

• -env, --environment

  • Name of the environment to run
  • PigeonEnv3Joints
  • PigeonRetinalEnv

• -dir, --snapshot_directory

  • Path to the snapshot directory within src/rlkit_ppo/data/

• -bs, --body_speed

  • Body speed of the pigeon model

• The resulting visualization is under <SNAPSHOT_DIRECTORY>/body_trajectory/