How yjy0625/equibot works
Main Components
• Simulation Environments: This component (located in equibot/envs/sim_mobile and related files) includes:
• Assets and URDF/mesh files for robots and objects (e.g., cloth folding, covering, closing).
• Environment definitions such as base_env.py, closing_env.py, covering_env.py, and folding_env.py.
• Utility scripts (e.g., bullet_robot.py, multi_camera.py, plan_control_traj.py) that support simulation, physics, and environment management.
• Scripts for demonstration generation (generate_demos.py) that produce input point clouds.
• Policy Module (Training & Evaluation): This module is found in equibot/policies/ and includes:
• Agent implementations and policies (dp_agent.py, equibot_agent.py, etc.).
• Configuration YAML files dictating training settings under equibot/policies/configs.
• Training, evaluation, and dataset code (train.py, eval.py, eval commands, etc.).
• Diffusion-related implementations under utils/diffusion and utils/equivariant_diffusion.
• Vision components (sim3_encoder.py, pointnet_encoder.py) for processing point clouds.
• External Dependencies & Infrastructure:
• Required libraries such as PyTorch, TorchVision, PyTorch3D.
• Integration with wandb for logging and experiment tracking.
• Build and installation configurations (setup.py, conda environment setup).
Relationships and Interactions
• Data Flow:
• Demonstration Generation: “generate_demos.py” produces demo data (point clouds) from various simulation environments.
• Training: The training scripts in policies use these demos; configuration files (YAML) specify behavior for each simulation setup.
• Evaluation: Evaluation scripts launch policies in simulated environments with varied parameters (e.g., R+Su, R+Sn, R+Sn+P) reflecting randomized conditions.
• Component Dependencies:
• The simulation environments depend on asset files (robots, meshes, textures) and utility scripts that define environment physics.
• The policy training and evaluation rely on the simulation outputs, configuration files, and deep learning frameworks.
• Modules in the “utils” subdirectories provide shared functionalities (e.g., coordinate transformations, frame conversions) used by both environments and policies.
Architectural Patterns and Design Principles
• Modularity: The clear separation between simulation (environment setup, data generation) and policy (training, evaluation, diffusion modules) underlines a modular design.
• Config-driven Behavior: The use of YAML configuration files indicates a design where behavior can be tuned externally without modifying core code.
• Data Pipeline: A pipeline flows from demonstration generation to training and evaluation, indicating a study of data-informed learning.