How warplab/isaac-auv-env works
Project Type and Purpose
• Type: Research-grade reinforcement-learning (RL) simulation environment and training framework.
• Purpose: Provide a Gym-style environment (WarpAUV) atop Omniverse IsaacSim/IsaacLab, run PPO training, evaluate policies, and manage assets/configurations.
File Structure
• Top level:
– agents/ ← RL algorithm configuration (rsl_rl_ppo_cfg.py)
– custom_workflows/ ← CLI entry points for train/eval/plot
– warpauv_env.py ← Gym-style environment wrapper
– rigid_body_hydrodynamics.py, thruster_dynamics.py ← custom physics submodules
– assets/warpauv.py ← AUV USD asset configuration
– data/warpauv/ ← USD meshes and configuration YAML
– weights/ ← Saved model checkpoints
• External dependencies:
– IsaacSim v4.5.0 + IsaacLab v2.2.0 (Omniverse GPU simulation engine)
– Gym, PyTorch (implied by .pt), NVIDIA RL utilities
Extract Key Architecture Notes from the README
• The simulation engine (IsaacSim) is decoupled via IsaacLab workflows.
• Training is driven by a CLI script (scripts/reinforcement_learning/rsl_rl/train.py).
• The environment exposes gym spaces (obs/action/state) and loads the USD asset.
Main Components
A. Reinforcement-Learning Agent
– Config (agents/rsl_rl_ppo_cfg.py)
– PPO algorithm (PyTorch)
– Checkpoint storage (weights/)
B. Gym-Style Environment Wrapper (warpauv_env.py)
– Defines observation_space, action_space, step/reset
– Integrates physics and thruster modules
C. Physics Submodules
– rigid_body_hydrodynamics.py
– thruster_dynamics.py
D. Asset Loader (assets/warpauv.py)
– Reads USD model from data/warpauv/warpauv.usd
– Configures RigidBodyProperties, Articulation props
E. IsaacSim/IsaacLab Simulation Engine
– External service
– Runs the real-time physics & rendering loop
F. CLI Workflows (custom_workflows/)
– train.py → kicks off agent + env + simulation
– play*.py, play_eval*.py → evaluation and data collection
– plot_metrics.py → post-training visualization
G. Data/Configuration
– USD meshes (instanceable_meshes.usd)
– config.yaml for asset parameters
Relationships & Data Flows
• Agent → [action] → Environment
• Environment → [step] → IsaacSim Engine
• IsaacSim Engine → [physics/render] → Physics Submodules & USD assets
• IsaacSim Engine → [obs, reward, done] → Environment → Agent
• Agent → [policy update] → Agent (self-contained training loop)
• Training/Eval CLI → orchestrates Agent + Environment instantiation
• Plot workflow → consumes metrics log output
Architectural Patterns & Design Principles
• Modular layering:
– External engine (IsaacSim) decoupled from Python wrapper
– Physics concerns split into dedicated modules
• Standard RL interfaces: Gym API for compatibility
• Configuration-driven: all hyperparams & asset settings live in code and YAML