Hero Image

Simulate,evaluate, and learn

A controllable and interactive simulation engine built for vision researchers.

What is LychSim

LychSim is a highly controllable, interactive simulation framework built on Unreal Engine 5, designed to lower the technical barrier of using a modern game engine for computer vision research.

Python and
Agentic Integration

One scene, three entry points. Script experiments through the Python API, hand the wheel to an LLM via MCP, or manage running instances from the CLI.

Streamlined Python API

CameraObjectDataMixins

A single LychSim class composes Camera, Object, and Data mixins. Capture RGB, depth, normals, segmentation, and point maps; query and manipulate actors; pause and resume the engine — all from native Python, no boilerplate.

from lychsim.api import LychSim
 
sim = LychSim(server_name="localhost", port=9000)
rgb    = sim.get_cam_lit(cam_id=0)
depth  = sim.get_cam_depth(cam_id=0)
annots = sim.get_obj_annots()

Key
Features

Key features and functionalities
of LychSim

2D3DVisibility
Comprehensive 2D and 3D ground truths for training, evaluation, rewards, etc.
Adversarial ExaminerEmbodied AISpatial Understanding
Interactive simulation for adversarial examiner, embodied AI, etc.
VehiclePedestrianRoadSidewalk
Procedural rules for controllable simulation
Multi-ViewOptimiation
Parallel rendering for fast simulation-in-the-loop

0+

Scenes annotated with procedural rules

Illustration

0+

Objects with category,
size, color, and material annotations

Illustration

Case Studies

How LychSim powers real research — from RL-driven model stress-testing to LLM-controlled scene composition.

Adversarial Examiners

RLRobustnessClosed-loop

Standard datasets cover only a narrow slice of the real-world parameter space. LychSim's controllable simulation lets an RL policy explore 3D camera viewpoints and scene configurations to surface vision-model failures — for example, minimizing Segment Anything's IoU on a target object exposes weaknesses on common objects in simple environments.

RL-based adversarial examiner exposes Segment Anything's failure modes by exploring 3D camera viewpoints around a target.

Interactive Scene Planning

LLMMCPScene planning

The MCP integration turns LychSim into a closed-loop playground for language-driven scene layout. Agentic LLMs query scene state, plan layouts from natural language, place actors, and verify the result — all through the same standardized tool calls.

Language-driven scene planning: an agent constructs and edits 3D layouts through MCP tool calls.

Public Release

We release the LychSim simulation framework, two Hugging Face datasets (object pose alignments and scene-level procedural rules), and the technical report describing the full system.

LychSim teaser figure: controllable, interactive simulation with rich 2D and 3D ground truth

LychSim

Simulate

Evaluate

Learn

LychSim

Simulate

Evaluate

Learn