Architecture Explorer

Explore EmbodiChain through a system overview and five specialist views: Task Program, Simulation, Data & Learning, Generation & Toolkits, and Toolkits. Select a module to inspect its responsibility, follow direct relationships, and read the source evidence behind each connection. Source links identify the exact commit being documented; documentation links stay within this documentation version.

The graph describes selected static relationships. It is not a runtime trace, an execution sequence, or a complete dependency inventory. Reviewed semantic relationships and extracted import declarations are labeled separately.

Each registered project-context topic and top-level production package has a module entry. This is an entry-point coverage measure: internal implementations and relationships are still selected, and an unconnected module may be consumed through an external host or a configuration-driven integration.

Open full-screen explorer

The module reference below is searchable and remains available without JavaScript. The explorer follows the documentation theme. Open full-screen to keep your current selection, or use Share view inside the explorer to copy a full-screen link with the current selection and filters.

Module reference

Source revision: 7ae9a5012091adce9d9e3e1acfad4e5dbf8bc318 · package 0.2.4

Curated static relationships; not a complete dependency graph or runtime trace.

Module	Responsibility	Evidence and documentation
CLI	Dispatches commands and lazily loads task discovery, environment, and training entry points.	Source · CLI reference
Environment launcher	Runs task discovery, configuration loading, and Gym environment creation.	Source · Environment launch guide
Task discovery	Imports installed task packages through entry points to populate the environment registry.	Source · Task discovery and supported tasks
Official tasks	Registers official task entry points and organises reusable robot tasks.	Source · Supported tasks
Config loader	Loads deployment files, resolves components, and assembles the typed environment configuration.	Source · Configuration guide
Component composition	Resolves environment and embodiment components into runnable Gym configurations.	Source · Configuration guide
EmbodiedEnv	Assembles scenes, managers, and demonstrations with optional Task Program integration.	Source · EmbodiedEnv documentation
BaseEnv	Owns the Gym loop, simulation initialisation, and environment step lifecycle.	Source · BaseEnv documentation
Managers	Organises observations, actions, rewards, events, and data recording within the environment lifecycle.	Source · Gym managers and environment lifecycle
Configuration utilities	Defines typed configuration copying, serialization, and required-field validation.	Source · Configuration utilities documentation
Domain randomization	Event functors randomize physical, visual, spatial, and geometric properties.	Source · Domain randomization documentation
DifferentiableEnv	Preserves the Gym lifecycle while bridging task-defined kinematics to autograd.	Source · DifferentiableEnv documentation
EventManager	Dispatches configured event functors with lifecycle scheduling and scoped seeds.	Source · EventManager documentation
TaskProgramCompiler	Compiles task configuration into an analysable program using a static scene manifest.	Source · Task Program guide · TaskProgramCompiler documentation
SemanticCallCompiler	Analyses and lowers Semantic Calls against the bound scene and skill integration.	Source · Task Program guide · SemanticCallCompiler documentation
SemanticCallExecutor	Lowers and executes semantic calls, creating an execution session for each call.	Source · Task Program guide · SemanticCallExecutor documentation
TaskProgramEnvironmentAdapter	Validates integration choices and assembles the semantic runtime and Gym bridge.	Source · Task Program guide · TaskProgramEnvironmentAdapter documentation
TaskProgramDemoBridge	Converts program segments into lazy Gym demonstrations with a shared runtime, buffer, and clock.	Source · Task Program guide · TaskProgramDemoBridge documentation
MLLM frontend	Decodes untrusted model JSON and compiles it through the existing Task Program adapter.	Source · MLLM frontend documentation
AtomicActionEngine	Manages planning resources and atomic skills to create executable action sessions.	Source · Atomic Skills guide · AtomicActionEngine documentation
MotionGenerator	Composes planning backends behind a unified robot motion generation interface.	Source · Motion generation guide · MotionGenerator documentation
Planners	Provides motion planning backends that generate constrained robot trajectories.	Source · Motion planner guide
Solvers	Provides inverse kinematics backends and stateful robot solving capabilities.	Source · Kinematic solver guide
Robot workspace	Loads reachability caches and samples robot joint configurations and poses.	Source · Robot workspace documentation
Trajectory augmentation	Tracks trajectory candidates, generation budgets, coverage, and commit receipts.	Source · Trajectory augmentation documentation
SimulationManager	Manages the world, robots, sensors, physics, and visualisation lifecycle.	Source · SimulationManager documentation
Robot	Represents a simulated robot with control parts, joint state, and kinematics.	Source · Robots and articulations
Sensors	Provides camera, stereo, and contact observations for the environment.	Source · Sensor guide
Physics backends	Constructs the configured physics backend and connects simulation state to physics solving.	Source · Physics backend guide
Browser visualization	Publishes simulation snapshots and interactive controls through a browser runtime.	Source · Browser visualization guide
Rigid objects	Represents rigid scene objects and their physical state.	Source · Rigid objects documentation
Rigid object groups	Groups multiple rigid bodies into a batched scene entity.	Source · Rigid object group guide · Object API reference
Articulations	Represents jointed scene entities and articulated state.	Source · Articulations documentation
Deformable objects	Represents Newton-backed deformable particle-set entities.	Source · Deformable objects documentation
Lights	Controls light color, intensity, falloff, and pose through a batched facade.	Source · Lights documentation
Rigid constraints	Wraps per-arena fixed constraints between two rigid objects.	Source · Rigid constraints documentation
Gizmo	Provides native robot IK and browser manipulation controls for simulation targets.	Source · Gizmo documentation
Compute	Shared numerical algorithms for kinematics, trajectories, geometry, and images.	Source · Compute API reference
Trajectory computation	Shared interpolation, resampling, timing, and trajectory warping algorithms.	Source · Trajectory computation documentation
Kinematics computation	Shared numerical kinematics and manipulability algorithms.	Source · Kinematics computation documentation
DatasetManager	Coordinates dataset functors for frame collection, saving, and finalization.	Source · DatasetManager documentation
LeRobot recording	Persists demonstrations and metadata in LeRobot datasets with optional depth sidecars.	Source · LeRobot recording documentation
Async recording	Moves durable dataset writes to one background worker.	Source · Async recording documentation
Depth sidecar writer	Coordinates depth-video files and metadata alongside recorded demonstrations.	Source · Depth video documentation
Depth sidecar reader	Loads depth metadata and decodes episode/sensor depth video frames.	Source · Depth sidecar reader documentation
Depth quantization	Converts depth values to and from the sidecar video representation.	Source · Depth quantization documentation
OnlineDataEngine	Online data engine managing workers and their lifecycle.	Source · Online data streaming guide
OnlineDataset	Exposes live engine samples as an infinite PyTorch IterableDataset.	Source · OnlineDataset documentation
Chunk samplers	Selects fixed-shape time windows using configurable chunk-length distributions.	Source · Chunk samplers documentation
Data assets	Resolves asset paths and coordinates registered dataset downloads and caches.	Source · Data assets documentation
Scene generation	Generates scene state from input images for scene asset workflows.	Source · Scene Engine guide
Scene understanding	Interprets an image, segments objects, and initializes a validated support graph.	Source · Scene Engine guide
Asset generation & refinement	Generates geometry and articulated assets, then refines scene placement.	Source · Scene Engine guide
Scene asset preparation	Prepares generated Scene Engine geometry for layout and export.	Source · Scene Engine guide
Scene editing	Edits an existing portable scene using a typed edit plan.	Source · Scene editing documentation
Scene edit layout	Reconstructs scene placement against an edited goal graph.	Source · Scene editing documentation
Scene import	Reconstructs an editable scene and graph from a portable scene export.	Source · Scene export documentation
Scene export	Writes portable scene metadata and assets for downstream simulation preparation.	Source · Scene export documentation
SimReady pipeline	Ingests raw assets through parsing and conversion stages into SimReady outputs.	Source · SimReady pipeline documentation
Asset parser stages	Dispatches inspector, geometry, physics, USD, and internal asset parsers.	Source · SimReady pipeline documentation
GenSim UI sessions	Tracks one replaceable subprocess per Gradio session.	Source · Scene Engine guide
Training assembly	Builds environments, policies, algorithms, and the compatible trainer from configuration.	Source · Training assembly documentation
RL algorithms	Registers PPO, GRPO, and APG with their rollout contracts.	Source · RL algorithms documentation
PPO	Optimizes an actor-critic policy from standard TensorDict rollouts.	Source · PPO documentation
GRPO	Optimizes policies with group-relative advantages from standard rollouts.	Source · GRPO documentation
APG	Optimizes policy parameters through differentiable rollout rewards.	Source · APG documentation
Policy models	Defines policy action inference and model interfaces.	Source · Policy models documentation
ActorCritic	Provides Gaussian actions and value estimates with separate actor and critic inputs.	Source · ActorCritic documentation
ActorOnly	Provides Gaussian actions without a learned value network.	Source · ActorOnly documentation
Learning environments	Defines lightweight vector-environment protocols and factory registration.	Source · Learning environments documentation
Standard trainer	Coordinates standard rollout collection, algorithm updates, evaluation, and checkpoints.	Source · Reinforcement learning overview
SyncCollector	Collects standard vector-environment rollouts into the training buffer.	Source · SyncCollector documentation
Rollout buffer	Stores standard rollout transitions and bootstrap state in TensorDict buffers.	Source · Rollout buffer documentation
DifferentiableTrainer	Coordinates graph-connected rollouts and analytic policy-gradient updates.	Source · DifferentiableTrainer documentation
Differentiable collector	Collects differentiable trajectories while preserving their autograd graph.	Source · Differentiable collector documentation
Running observation normalization	Normalizes continuous observation fields with explicit running-statistic updates.	Source · Running observation normalization documentation
Policy evaluation	Runs trained policy evaluation in headless or viewer workflows.	Source · Policy evaluation documentation
Devices	Defines start, stop, and controller-state interfaces for robot input devices.	Source · Laboratory and device API reference
Device controller	Maps input-device state to robot joint commands with joint-limit handling.	Source · Device controller documentation
GraspKit	Defines parallel-jaw gripper models and grasp-pose generation interfaces.	Source · Grasp generation documentation
URDF assembly	Assembles URDF components, connections, sensors, and mesh resources.	Source · URDF assembly documentation
Convex decomposition	Prepares convex collision meshes for URDF assets.	Source · Convex decomposition documentation

Additional modules in specialist views

Module	Responsibility	Evidence and documentation
Antipodal grasp generator	Generates candidate grasp poses through a reusable parallel-jaw interface.	Source · Grasp generation documentation
Newton autograd bridge	Connects Warp tape derivatives to PyTorch action and functional-state gradients.	Source · Newton autograd bridge documentation
BatchEntity	Defines shared configuration and instance operations for batched simulation entities.	Source · BatchEntity documentation
URDF collision processor	Replaces mesh collisions with convex pieces and optionally recomputes inertia.	Source · Convex decomposition documentation
BufferedGymCommandSink	Validates and buffers control commands for incremental Gym demonstration consumption.	Source · Task Program guide · BufferedGymCommandSink documentation
CompiledTaskProgram	Immutable compilation result with materialized task segments and analysis views.	Source · Task Program guide · CompiledTaskProgram documentation
Effect evidence	Collects typed evidence from registered providers for semantic effects.	Source · Effect evidence documentation
ExecutionRunner	Advances atomic skill sessions through observation, command, and clock interfaces.	Source · Atomic Skills execution guide · ExecutionRunner documentation
Interactive grasp annotation	Offers browser selection of a graspable mesh region.	Source · Grasp generation documentation
Mesh grasp backend	Coordinates sampling, annotation, collision checks, and caches for one mesh.	Source · Grasp generation documentation
Gripper collision checks	Checks gripper geometry against the target mesh.	Source · Grasp generation documentation
Antipodal sampling	Samples opposing surface contacts using mesh ray casting.	Source · Grasp generation documentation
Parallel-jaw gripper model	Defines physical gripper dimensions independently of the sampling algorithm.	Source · Grasp generation documentation
Integration catalog	Describes one Task Program integration and validates compilation against its fingerprint.	Source · Integration catalog documentation
Object backend views	Adapts rigid and articulated object state to backend-neutral Scene batches.	Source · Object backend views documentation
Parallel execution	Coordinates parallel semantic branches through command and safety interfaces.	Source · Parallel execution documentation
Program loading & validation	Loads bounded JSON or YAML into the canonical Task Program configuration.	Source · Program loading & validation documentation
SceneRegistry	Binds semantic entity identifiers to scene metadata and runtime providers.	Source · SceneRegistry documentation
SimulationTaskProgramFactory	Binds simulated robots, planners, and services to construct the atomic skill engine.	Source · Task Program guide · Simulation integration API
RobotSkillProfile	Declares robot resources, skill bindings, and effect-assurance capabilities.	Source · RobotSkillProfile documentation
EnvironmentStepClock	Maintains execution time on the environment step grid.	Source · Task Program guide · Gym Task Program API
Surface deformables	Specializes deformable objects for cloth particle sets.	Source · Surface deformables documentation
TaskProgramCfg	Declarative task configuration describing intent and execution structure.	Source · Task Program guide · TaskProgramCfg documentation
Toolkit directory	Standalone grasp sampling, collision-mesh preparation, and URDF assembly utilities.	Source · Toolkits documentation
URDF components	Processes component links, joints, naming, and mesh references.	Source · URDF assembly documentation
URDF connections	Applies component connection rules and sensor attachment joints.	Source · URDF assembly documentation
URDF mesh resources	Copies mesh resources and rewrites references for the assembled robot.	Source · URDF assembly documentation
URDF sensor attachments	Processes sensor descriptions and attachment metadata for assembly.	Source · URDF assembly documentation
URDF file output	Formats and persists the assembled URDF document.	Source · URDF assembly documentation
Volume deformables	Specializes deformable objects for volumetric soft-body particle sets.	Source · Volume deformables documentation
Workspace analyzer	Analyzes robot reachability and metrics for reusable workspace caches.	Source · Workspace analyzer documentation

Coverage and interpretation

Snapshot: 113 nodes, 113 source-reviewed relationships, and 114 statically extracted relationships across all views.

Import declarations do not establish runtime calls or execution order. Missing edges do not imply architectural independence.

Topics without represented nodes: none. Topic representation does not imply complete coverage.

Overview nodes without mapped relationships: Trajectory augmentation, Kinematics computation, GenSim UI sessions, URDF assembly.

• A curated source snapshot, not a complete import graph, runtime trace, or execution order.

• All registered context topics and top-level production packages have an entry point. Internal classes, algorithms, runtime paths, and relationships remain selectively represented.

• References through protocols and injected factories are scoped to the described production assembly.

• Source-reviewed means checked against the pinned source revision; it is not a physical qualification result.

• Static extraction scans the selected node source files only. Topic or package representation is not a completeness claim.