Add sim agent tutorial

Author: daphne-cornelisse

baselines/ppo/config/ppo_base_puffer.yaml

@@ -2,7 +2,7 @@ mode: "train"
 use_rnn: false
 eval_model_path: null
 baseline: false
-data_dir: /scratch/kj2676/gpudrive/data/processed/training
+data_dir: data/processed/training
 continue_training: false
 model_cpt: null
 
@@ -17,7 +17,7 @@ environment: # Overrides default environment configs (see pygpudrive/env/config.
   norm_obs: true
   remove_non_vehicles: true # If false, all agents are included (vehicles, pedestrians, cyclists)
   lidar_obs: false # NOTE: Setting this to true currently turns of the other observation types
-  reward_type: "weighted_combination" #"distance_to_logs"
+  reward_type: "weighted_combination" 
   collision_weight: -0.75
   off_road_weight: -0.75
   goal_achieved_weight: 1.0
@@ -67,7 +67,7 @@ train:
   clip_vloss: false
   vf_clip_coef: 0.2
   ent_coef: 0.0001
-  vf_coef: 0.5
+  vf_coef: 0.3
   max_grad_norm: 0.5
   target_kl: null
   log_window: 1000
@@ -85,11 +85,11 @@ train:
   checkpoint_path: "./runs"
 
   # # # Rendering # # #
-  render: True # Determines whether to render the environment (note: will slow down training)
-  render_interval: 10 # Render every k iterations
-  render_k_scenarios: 15 # Number of scenarios to render
-  render_3d: True
-  render_format: "gif" # Options: gif, mp4
+  render: false # Determines whether to render the environment (note: will slow down training)
+  render_3d: true # Render simulator state in 3d or 2d
+  render_interval: 1 # Render every k iterations
+  render_k_scenarios: 10 # Number of scenarios to render
+  render_format: "mp4" # Options: gif, mp4
   render_fps: 15 # Frames per second
   zoom_radius: 50
 

data/processed/examples/tfrecord-00000-of-01000_222.json

@@ -2,10 +2,10 @@ res_path: examples/experimental/eval/dataframes # Store dataframes here
 test_dataset_size: 10_000 # Number of test scenarios to evaluate on
 
 # Environment settings
-train_dir: /scratch/kj2676/gpudrive/data/processed/training
-test_dir: /scratch/kj2676/gpudrive/data/processed/validation
+train_dir: data/processed/training 
+test_dir: data/processed/validation 
 
-num_worlds: 2 # Number of parallel environments for evaluation
+num_worlds: 50 # Number of parallel environments for evaluation
 max_controlled_agents: 64 # Maximum number of agents controlled by the model.
 ego_state: true
 road_map_obs: true

data/processed/examples/tfrecord-00000-of-01000_4.json

@@ -0,0 +1,21 @@
+# Overrides default environment configs (see pygpudrive/env/config.py)
+max_controlled_agents: 64 # Maximum number of agents controlled by the model. Make sure this aligns with the variable kMaxAgentCount in src/consts.hpp
+ego_state: true
+road_map_obs: true
+partner_obs: true
+norm_obs: true
+remove_non_vehicles: true # If false, all agents are included (vehicles, pedestrians, cyclists)
+lidar_obs: false # NOTE: Setting this to true currently turns of the other observation types
+reward_type: "weighted_combination" 
+collision_weight: -0.75
+off_road_weight: -0.75
+goal_achieved_weight: 1.0
+dynamics_model: "classic"
+collision_behavior: "ignore" # Options: "remove", "stop", "ignore"
+dist_to_goal_threshold: 2.0
+polyline_reduction_threshold: 0.1 # Rate at which to sample points from the polyline (0 is use all closest points, 1 maximum sparsity), needs to be balanced with kMaxAgentMapObservationsCount
+sampling_seed: 42 # If given, the set of scenes to sample from will be deterministic, if None, the set of scenes will be random
+obs_radius: 50.0 # Visibility radius of the agents
+action_space_steer_disc: 13
+action_space_accel_disc: 7
+init_mode: all_non_trivial

examples/experimental/eval/config/eval_config.yaml

@@ -5,8 +5,6 @@
 from box import Box
 import numpy as np
 import dataclasses
-import os
-from pathlib import Path
 
 from gpudrive.env.config import EnvConfig, RenderConfig
 from gpudrive.env.env_torch import GPUDriveTorchEnv
@@ -18,7 +16,6 @@
 
 import logging
 import torch
-import random
 
 logging.basicConfig(level=logging.INFO)