Bugs related to the SDC in scenarios

[vcharraut]

Hello! I've encountered two issues while implementing RL training using Waymax, both related to the SDC (Self-Driving Car) simulations and dataset configurations.

1. SDC Presence with Low max_num_objects

Expected Behavior: When running simulations with a max_num_objects parameter, I expect the SDC to be present in the scene according to the specified object limit.

Actual Behavior: With max_num_objects set to 8, there are instances where the SDC does not appear in the scene at all. This results in the policy receiving the maximum reward without any SDC-related actions. Moreover, with max_num_objects set to 1, the SDC is almost never present in the scenarios.

Is it possible to force the presence of a SDC in any scenario ?

2. Error When Including SDC Paths

Expected Behavior: When setting include_sdc_paths=True in the dataset configuration, I expect the SDC paths to be included without any errors.

Actual Behavior: Upon setting include_sdc_paths=True, the following error is encountered:

Code:

scenarios = dataloader.simulator_state_generator(
    dataclasses.replace(
        config.WOD_1_1_0_TRAINING,
        max_num_objects=max_num_objects,
        batch_dims=(num_envs,),
        distributed=True,
        include_sdc_paths=True,
        num_paths=10,
        num_points_per_path=100,
    ),
)

Result:

tensorflow.python.framework.errors_impl.InvalidArgumentError: {{function_node __wrapped__IteratorGetNext_output_types_36_device_/job:localhost/replica:0/task:0/device:CPU:0}} Feature: path_samples/arc_length (data type: float) is required but could not be found.