How autonobot-lab/bestman works
Project Type & Purpose
• BestMan is a robotics platform for mobile manipulation with a focus on sim-to-real integration and modular algorithm deployment.
• Its purpose is to simulate and control a mobile manipulator (wheel-base with an arm) using a rich set of components such as digital twins, motion planning algorithms, perception modules, and a unified Robotics API.
• The platform provides both simulation (through PyBullet) and real-world hardware integration with demo examples and Blender-based rendering.
Key Components
a. Simulation & Environment:
• PyBullet Simulation Engine (core of simulation) – the “Env” directory and related Config YAML files.
• DigitalTwin modules manage digital replicas of hardware components.
b. Perception:
• Multiple perception pipelines for object detection (OWL_ViT), grasp pose estimation (AnyGrasp, Contact_GraspNet), affordance recognition, and segmentation (Detic, Grounded SAM, Lang SAM, Segment Anything).
• These are organized under the “Perception” folder.
c. Motion Planning:
• Manipulation and Navigation planning – include submodules like OMPL_Planner, A_star, PRM, and RRT.
• Incorporate Collision Detection modules (both “Basic_CD” and “FCL” versions).
d. Control & Sensor:
• PIDController under “Controller”.
• Sensor modules (such as Camera in “Sensor”).
e. Robotics API:
• Unified interfaces to interact with simulation and hardware (in “Robotics_API”).
f. Task Planning:
• Contains task-level and high-level planning (e.g., LLM_P and potentially PDDL streaming components).
g. User Interface & Visualization:
• User_Interface as integration layer for user interactions.
• Visualization modules handling simulation rendering, including the pyBullet-Blender integration (both under Visualization and docs for rendering integration).
h. Supporting Infrastructure:
• Config (various YAML files for different simulation and hardware scenarios).
• Examples that show individual demo setups.
• Install scripts and documentation to support setup and usage.
Relationships & Interactions
• The Simulation Engine and DigitalTwin (bidirectional data update: simulation state ↔ digital hardware models).
• Sensors feeding data into Perception modules; perception outputs then guide Motion Planning.
• Motion Planning sending commands to the Controller (PID) and then to the Robotics API to effect actions.
• Task Planning and High-Level Decision-making modules interfacing with both the motion and perception modules.
• Visualization and Blender rendering modules receiving simulation state and simulation logs to produce rendered outputs.
• Integration paths between the different modules (e.g., how Config parameters are universally used by Simulation, Motion Planning, and Perception modules).
Architectural Patterns & Design Principles
• Modularity: each major component (simulation, planning, perception, control) is a self-contained module sitting behind a unified API.
• Layer separation: Input (sensors, UI) → Processing (perception, planning, control) → Actuation (Robotics API, simulation actuation).
• Inter-module communication is predominantly via defined configuration interfaces (YAML files in Config) and standardized Python APIs.
Technologies & Frameworks
• PyBullet for simulation.
• Python for scripting with a variety of libraries (e.g., OMPL for planning, pre-commit setup using Black and isort for code quality).
• Blender, via the customized pyBulletSimImporter, for high fidelity rendering.
• ROS-like system abstractions (though not explicitly ROS, the layered architecture is similar to common robotics frameworks).