Using IPPO in GPUDrive

PufferLib implementation

Dependencies

This implementation is compatible with the gpudrive branch of PufferLib. To install, run:

pip install git+https://github.com/PufferAI/PufferLib.git@gpudrive

Example

Launch a run:

python baselines/ippo/ippo_pufferlib.py

Configs are in baselines/ippo/config/ippo_ff_puffer.yaml
A small feedforward network is implemented in integrations/rl/puffer/utils.py

Stable baselines 3 implementation

Example

Launch a run:

python baselines/ippo/ippo_sb3.py

Configurations are found in baselines/ippo/config/ippo_ff_sb3.yaml

Details

Resampling the data

The configuration for resampling traffic scenarios includes:

resample_scenarios: A boolean that enables or disables traffic scenario resampling when set to True.
resample_criterion: Set to "global_step", indicating resampling occurs based on the global step count.
resample_freq: Specifies resampling frequency at 50,000 steps, recommended to align with num_worlds * n_steps.
resample_mode: Set to "random" for random selection of new scenarios.

# RESAMPLE TRAFFIC SCENARIOS
resample_scenarios: bool = True
resample_criterion: str = "global_step" # Options: "global_step"
resample_freq: int = 100_000 # Resample every k steps (recommended to be a multiple of num_worlds * n_steps)
resample_mode: str = "random" # Options: "random"

Implemented networks

Classic Observations

For classic observations (e.g., ego_state), there is support for a permutation equivariant network (recommended). In baselines/ippo/config.py, set the following:

# NETWORK
mlp_class = LateFusionNet
policy = LateFusionPolicy

The default settings for classic observations are:

ego_state: bool = True  # Use ego vehicle state
road_map_obs: bool = True  # Use road graph data
partner_obs: bool = True  # Include partner vehicle information
norm_obs: bool = True  # Normalize observations

LiDAR Observations

For only LiDAR-based observations, set the following options:

ego_state: bool = False  # Use ego vehicle state
road_map_obs: bool = False  # Use road graph data
partner_obs: bool = False  # Include partner vehicle information
norm_obs: bool = False  # Normalize observations
disable_classic_obs: bool = True  # Disable classic observations for faster sim
lidar_obs: bool = True  # Use LiDAR in observations

You can also mix classic and LiDAR observations by setting:

ego_state: bool = True  # Include ego vehicle state in observations
road_map_obs: bool = True  # Include road graph in observations
partner_obs: bool = True  # Include partner vehicle info in observations
norm_obs: bool = True  # Normalize observations
disable_classic_obs: bool = False  # Keep classic observations
lidar_obs: bool = True  # Add LiDAR to observations

In both cases, you can use a feedforward network from networks/basic_ffn.py:

# NETWORK
mlp_class = FFN
policy = FeedForwardPolicy