Goal Radius Randomization

[riccardosavorgnan]

Goal radius randomization. Based on the work by Kevin; remaking this branch was simpler than rebasing the other branch after all the merge conflicts we had.

• first commit
• Working version with rendering working as well
• commit with config comments

[greptile-apps]

Greptile Summary

This PR adds per-goal randomization of the goal radius for each agent. Previously, the goal radius was a fixed global value (env->goal_radius). Now, when reward_randomization is enabled, each time a new goal is sampled (via sample_new_goal), the goal radius is independently randomized from a uniform distribution over [reward_bound_goal_radius_min, reward_bound_goal_radius_max] (configured as 2.0–12.0m). When randomization is disabled, the fixed goal_radius value is used as before.

• New sample_new_goal_radius function added and called at the end of sample_new_goal to decouple goal radius sampling from the rest of reward coefficient generation
• Distance check in c_step and all three rendering paths updated to use the per-agent reward_coefs[REWARD_COEF_GOAL_RADIUS] instead of the global env->goal_radius
• Goal radius assignment is duplicated between generate_reward_coefs (called at init/reset) and the new sample_new_goal_radius (called per new goal) — consider consolidating to reduce redundancy
• Config comment updated to note that the fixed goal_radius value is only active when reward_randomization = 0

Confidence Score: 4/5

• This PR is safe to merge — the logic is correct and the changes are well-scoped to goal radius randomization.
• Score of 4 reflects that the core logic is correct and consistent across all code paths (step logic, rendering modes). The only concerns are stylistic: duplicated goal radius assignment between two functions and mixed comment syntax in the INI file. No functional bugs or runtime risks identified.
• pufferlib/ocean/drive/drive.h — verify that the duplicated goal radius logic between generate_reward_coefs and sample_new_goal_radius behaves as intended across all goal behavior modes.

Important Files Changed

Filename	Overview
pufferlib/config/ocean/drive.ini	Updated comment for goal_radius to clarify it's only active when reward_randomization = 0. Minor mixed comment syntax style issue (; and //).
pufferlib/ocean/drive/drive.h	Adds sample_new_goal_radius function for per-goal radius randomization, updates distance check and all rendering paths to use per-agent radius. Logic is correct but introduces duplication with generate_reward_coefs.

Flowchart

flowchart TD
    A[Agent Init / Reset] -->|"generate_reward_coefs()"| B[Set all reward coefs including goal_radius]
    B --> C{goal_behavior?}
    C -->|GOAL_GENERATE_NEW| D[Agent reaches goal]
    D --> E["sample_new_goal()"]
    E --> F["sample_new_goal_radius()"]
    F --> G{reward_randomization?}
    G -->|Yes| H["random_uniform(min, max)"]
    G -->|No| I["env->goal_radius (fixed)"]
    H --> J[Set agent reward_coefs GOAL_RADIUS]
    I --> J
    J --> K["within_distance = distance < agent.reward_coefs[GOAL_RADIUS]"]
    C -->|GOAL_RESPAWN| L[Agent respawns]
    L -->|"generate_reward_coefs()"| M[Goal radius set at respawn]
    M --> K
    C -->|GOAL_STOP| N[Agent stops at goal]
    N --> K

Loading

_{Last reviewed commit: b73b835}

[greptile-apps]

_{2 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

Duplicated goal radius logic with generate_reward_coefs
generate_reward_coefs (line 417-418 and 449) already sets REWARD_COEF_GOAL_RADIUS using the same logic. When goal_behavior == GOAL_GENERATE_NEW, the value set by generate_reward_coefs at init/reset is immediately overwritten by this function when sample_new_goal is first called.

This isn't a bug — the separation is intentional so the goal radius can be re-randomized per goal. But you may want to consider removing the REWARD_COEF_GOAL_RADIUS assignment from generate_reward_coefs and instead calling sample_new_goal_radius directly after generate_reward_coefs in the init/reset paths. This would make sample_new_goal_radius the single source of truth for goal radius assignment, reducing duplication.

_{Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!}

[mpragnay]

lgtm!