How vita-epfl/crowdnav works
Project Type and Purpose
• This is a research‐oriented deep reinforcement learning framework for crowd‐aware robot navigation (no GUI/web front-end).
• It comprises two main packages:
• crowd_sim: a Gym-style crowd simulation environment (with RVO2 dynamics)
• crowd_nav: training/testing code for multiple RL policies (CADRL, LSTM-RL, SARL, etc.)
High-level Components
a. Simulation Environment (crowd_sim/)
• envs/crowd_sim.py: core simulator stepping through time
• envs/policy/*.py: built-in human navigation policies (e.g., ORCA, linear)
• envs/utils: state/action/human/robot abstractions & helper functions
b. Policy Implementations (crowd_nav/policy/)
• cadrl.py, lstm_rl.py, sarl.py, multi_human_rl.py: RL policy classes
• policy_factory.py: Factory pattern to instantiate policies by name
c. Configuration Loader (crowd_nav/configs/)
• env.config: environment hyperparams (crowd size, noise, time step)
• policy.config: network/policy hyperparams
• train.config: training loop settings (learning rate, batch size, episodes)
d. Training & Evaluation Scripts (crowd_nav/)
• train.py: orchestrates RL training
e. Utility Modules (crowd_nav/utils/)
• explorer.py: handles exploration strategy (ε-greedy)
• memory.py: replay buffer management
• trainer.py: main RL loop, backpropagation, logging
• plot.py: post-hoc learning curve visualization
Interactions and Data Flow
1. train.py
• Loads train.config, policy.config, env.config
• Instantiates Environment via crowd_sim API and Policy via policy_factory
• Instantiates Memory & Explorer
• Calls Trainer.run_training():
• for each episode:
– Reset env → initial state
– Loop over timesteps:
• Policy.compute_action(state) → action
• env.step(action) → next_state, reward, done
• Memory.store(transition)
• Explorer.adjust_epsilon()
• Trainer.learn_from(memory) (samples batch, updates policy network)
• Logs metrics to output.log
2. test.py
• Similar but no learning: loads a saved model, runs episodes, optionally visualizes trajectories
3. plot.py
• Reads output.log → plots loss/return curves
Key Architectural Patterns & Design Principles
• Modular separation: simulation vs. learning vs. config vs. policies
• Factory pattern to select policies by string key
• Config‐driven behavior without code changes
• Use of standard Gym API for env.step/reset
• Replay buffer decoupled from main loop for easy swapping (e.g., prioritized replay)