Remove vehicles with invalid positions and invalid goals from controlled vehicles

[daphne-cornelisse]

Problem description

In the env.reset() function, vehicles in invalid positions are removed from the scene but are still added to self.controlled_vehicles. This leads to invalid samples (invalid observations) that cause three issues:

1. We attempt to control vehicles that remain in invalid states, which happens every time env.step(.) is called.
2. An extra vehicle is counted when calculating the goal rate.
3. Invalid observations have an unrealistically large speed and distance to goal value, which introduces noise in the training data.

NOTE: these are vehicles that are not included in the valid_files.json, and so they are not set to be expert-controlled.

This fix (details below) reduces the number of invalid samples from 4.62% to 0.34% (the total number of scenes is 1000). The variable "self.invalid_samples" tracks the number of invalid samples.

Steps to reproduce

1. Download the nocturne mini dataset
2. Set the data_path: ./data_full/train in env_config
3. Comment out the fix
4. Run the code below
5. This should show the logs below

Code:

# Load environment variables and config
env_config = load_config("env_config")
  
# Initialize an environment
env = BaseEnv(config=env_config)

# This is an example scene that can be used to reproduce this error 
filename = "tfrecord-00703-of-01000_46.json"

# Reset to this specific scene
obs_dict = env.reset(filename)

# Step
env.step({})

Logs

DEBUG:root:(IN RESET) selecting vehicles to control, current list: []
DEBUG:root:looking at veh_id 26...
DEBUG:root:updated self.controlled_vehicles: [26]
DEBUG:root:looking at veh_id 17...
DEBUG:root:updated self.controlled_vehicles: [26, 17]
DEBUG:root:looking at veh_id 3...
DEBUG:root:veh_id 3 is INVALID!
DEBUG:root:updated self.controlled_vehicles: [26, 17, 3]

INFO:root:(IN STEP) t = 1 | tfrecord-00703-of-01000_46.json
INFO:root:veh_id = 3 | pos: -10000.0
INFO:root:controlled_vehs: [26, 17, 3] 

Solution

Added booleans to make sure vehicles at invalid positions and with invalid goals are not included in self.controlled_vehicles

# Otherwise the vehicle is valid and we add it to the list of controlled vehicles
if not vehicle.expert_control \
and not veh_at_invalid_pos \
and not veh_has_invalid_goal_pos \
and curr_index < self.config.max_num_vehicles:
	self.controlled_vehicles.append(vehicle)
	logging.debug(f'updated self.controlled_vehicles: {[veh.id for veh in self.controlled_vehicles]}')
	curr_index += 1

Questions

• Check: Would you mind reviewing this and checking if it does not introduce any new bugs?
• Open question: 0.34% of the samples are still invalid (with 1000 scenes, so it could increase/decrease with the full data set). This issue occurs in the env.step() function, where a vehicle's position is sometimes invalid (the position is -10,000), but then becomes normal again in the next step. Do you have any idea where this issue could be coming from?

[eugenevinitsky]

I'm guessing that this condition, which is post reset, is only getting triggered for vehicles that are expert controlled. The expert controlled vehicles will occasionally pop to the invalid position