Guidance, Rendering and Type-Aware Agents/Off-roads

[nadarenator]

No description provided.

[greptile-apps]

Greptile Overview

Greptile Summary

This PR implements guidance and type-aware agents for autonomous driving simulation. The main changes include:

• Guidance Waypoint Observations: Adds optional egocentric waypoint observations (182 features for 91 waypoints) to provide agents with trajectory guidance information from expert demonstrations
• Type-Aware Observations: Adds agent type (vehicle/pedestrian/cyclist) to ego observations, normalized by dividing by 3.0, enabling policies to condition behavior on agent type
• Dynamic Observation Space: Refactors observation size calculations throughout the codebase to be computed dynamically based on dynamics_model (classic vs jerk) and use_guidance_observations flag
• Evaluation Pipeline Updates: Major refactor of eval_pipeline.py to support evaluating policies trained with different dynamics models (classic vs jerk) separately
• Configuration Enhancements: Expands sweep configuration with additional hyperparameters for guidance reward tuning (guidance_speed_weight, guidance_heading_weight, waypoint_reach_threshold)
• Safety Improvements: Adds bounds check env->active_agent_count < env->num_agents in set_active_agents() and restricts road edge collision checks to vehicles only with agent->type == VEHICLE
• Human Data Handling: Makes human action likelihood loss conditional on human_sequences > 0 to prevent errors when human demonstration data is unavailable

Confidence Score: 4/5

• This PR is safe to merge with minor testing recommended for new observation features
• The implementation is well-structured with consistent changes across C/Python boundaries. The dynamic observation sizing is correctly propagated through all layers (drive.h, binding.c, drive.py, torch.py, drivenet.h, env_binding.h). Safety improvements (bounds checking, type-aware collision detection) reduce potential bugs. The optional nature of guidance observations (default=0) makes this backwards compatible. Score is 4 instead of 5 due to: (1) large observation space expansion (182 additional features) needs validation that it doesn't cause memory issues or training instability, and (2) the type normalization factor of 3.0 assumes max type value is 3 (CYCLIST) but should be verified against actual type enum values.
• Pay close attention to pufferlib/ocean/drive/drive.h for the new guidance observation logic and memory allocation changes

Important Files Changed

Filename	Overview
pufferlib/ocean/drive/drive.h	Adds type-aware agent observations and guidance waypoint observations. Includes safety fix for road edge collision checks and bounds checking for active agents.
pufferlib/ocean/drive/binding.c	Refactors configuration override system using macros for cleaner code. All settings now configurable via kwargs with INI defaults.
pufferlib/ocean/drive/drive.py	Adds use_guidance_observations parameter and properly calculates ego observation size including guidance features.
pufferlib/config/ocean/drive.ini	Adds use_guidance_observations config and expands sweep configuration with more hyperparameters for guidance reward tuning.
examples/eval_pipeline.py	Major refactor supporting multiple dynamics models (classic/jerk) with policy configurations and improved evaluation flow.
pufferlib/pufferl.py	Makes human action likelihood loss optional based on human_sequences config. Prevents errors when human data unavailable.
pufferlib/utils.py	Passes use_guidance_observations and dynamics_model to evaluation subprocesses for observation space consistency.
pufferlib/ocean/torch.py	Uses dynamic ego dimension from environment instead of hardcoded value, supporting variable observation sizes.

Sequence Diagram

sequenceDiagram
    participant Config as drive.ini Config
    participant Python as drive.py
    participant Binding as binding.c
    participant Core as drive.h
    participant Network as drivenet.h
    
    Config->>Binding: Load INI with use_guidance_observations
    Python->>Binding: Initialize env with guidance params
    Binding->>Core: Set guidance config + dynamics_model
    
    alt use_guidance_observations=1
        Core->>Core: ego_dim = base + GUIDANCE_OBS_SIZE (182)
    else use_guidance_observations=0
        Core->>Core: ego_dim = base only (8 or 11)
    end
    
    Core->>Core: compute_observations()
    Core->>Core: Add type feature to ego obs
    
    alt use_guidance_observations=1
        Core->>Core: Transform 91 waypoints to egocentric
        Core->>Core: Append to ego observations
    end
    
    Core->>Python: Return observations with correct size
    Python->>Network: Policy forward with dynamic ego_dim
    Network->>Python: Return actions
    Python->>Core: Step environment with actions
    
    alt use_guided_autonomy=1
        Core->>Core: Compute guidance reward
        Core->>Python: Return reward + guidance bonus
    end

Loading

[greptile-apps]

_{8 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

consider using a named constant like MAX_AGENT_TYPE instead of the hardcoded 3.0f for type normalization to improve maintainability

_{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!}