System architecture

SkyEngine separates the experiment workflow from factory execution and algorithm decisions. This lets different algorithms operate in a shared simulation environment.

User interface
Factory management · Live monitoring · Analysis · Experiments
                           │
Platform services
Run control · Experiment orchestration · Models · Results
                           │
Factory simulation
Jobs · Machines · AGVs · Roads · Materials · Disturbances
                           ↕
Algorithm components
Process scheduling · Task assignment · Path planning · Joint policies

Platform and factory lifecycle

The Vue interface talks to a FastAPI backend. Factory proxies provide lifecycle operations such as initialization, start, pause, reset, and cleanup, together with state, metric, and control streams. The backend selects a proxy for the environment being used.

The containerized grid environment uses DockerProxy to connect to simulation and algorithm services. PacketFactory uses a separate proxy and configuration flow. Shared lifecycle methods do not make the underlying environments interchangeable.

Simulation and decisions

In the grid environment, the coordinator connects three responsibilities:

Component	Responsibility
Job solver	Decide processing operations and machine assignments
Task assigner	Assign transport tasks to vehicles
Route solver	Decide vehicle movement on the road network

Joint scheduling policies can coordinate production and transport together. The DFJSP-T policy includes PIBT for vehicle movement; users do not select an additional MAPF algorithm for its training and model tests.

The factory remains responsible for legal movement, processing progression, buffer capacity, material handoffs, and completion. A final operation finishing is not sufficient: the job completes after delivery and unloading at the finished-goods station.

Experiments and evidence

The experiment workbench organizes training, parameter exploration, and testing. Models and checkpoints connect training to separate tests. Recorded events, dataset results, and replay help explain what happened during execution.

Use independent training, validation/tuning, and final benchmark datasets. A comparison should keep instances, seeds, disturbances, and simulation limits consistent.

Repository boundaries

Repository	Responsibility
skyengine	Platform, frontend, simulation, and experiment infrastructure
SkyEngine-FJSP	Process scheduling algorithms and HTTP container services
SkyEngine-MAPF	Multi-agent path planning algorithms and HTTP container services
skyengine-DFJSPT	The dfjsp_t_rl Python package and joint scheduling training components

Obtain the separate skyengine-DFJSPT package from the project maintainers and prepare it before installation.