01
Generate worlds
Language and diffusion models turn a text prompt or a 2D floorplan into a furnished, human-centric 3D world, drawing on a semantically annotated model database.
The platform
The whole research loop in one stack.
Modules are abstracted across simulators, so a scenario you design once can run in 2D for training and in photorealistic 3D for evaluation.
02
Simulate people
Swap between six pedestrian drivers, from social force to learned models, with groups, queues and need-driven behavior trees.
03
Train planners
Train DRL agents with Stable-Baselines3 from a single config file, with multiprocessed rollouts, curricula and faster-than-real-time 2D simulation.
04
Benchmark
One click runs planners across worlds and scenarios and records collisions, path length and time to goal, plus social metrics such as time in personal space.
