Curriculum Learning

URBAN-SIM supports curriculum learning by gradually adjusting task difficulty during training. Each curriculum behavior is defined as a CurrTerm function and bound via CurriculumCfg.

A common use case is to expand the movement range or goal distance as training progresses.

Example: Increasing Goal Distance

The following curriculum function is defined for COCO:

def increase_moving_distance(env, env_ids, command_name='pose_command', ...):
    cur_iteration = env.common_step_counter // num_steps_per_iteration
    x_left = start_x + (end_x - start_x) * (cur_iteration / total_iterations)
    env.command_manager.get_term(command_name).cfg.ranges.pos_x = (10., min(x_left, end_x))
    env.command_manager.get_term(command_name).cfg.ranges.pos_y = (0., map_region)

This function progressively expands the allowed x-axis goal region as training proceeds.

Curriculum Binding

It is registered via:

@configclass
class CurriculumCfg:
    increased_moving_distance = CurrTerm(
        func=increase_moving_distance,
        params={}
    )

This term will be called automatically during each environment step to update internal task parameters (e.g., goal position sampling bounds).

Design Notes

• Curriculum functions operate outside the RL policy loop and are called by the environment manager.

• You can define any callable that modifies task-related config values (goal ranges, difficulty level, distractor count, etc.).

• These changes take effect in the next episode reset.