github.devotd.
NewRobotics

omnilink-tech/omnisim — A Robotics Simulator for Agents

A look at omnilink-tech/omnisim's agent control features and the validation scope of its own benchmark.

Repositoryomnilink-tech/omnisim ↗

Open-source robotics simulator for coding agents: HTTP/JSON + MCP control, Newton physics, wgpu rendering, ROS 2, and reproducible benchmarks.

stars
★ 188
Forks
10
Language
Python
License
Apache-2.0
Latest release
v9.2.6
Read original source ↗View on GitHub ↗

omnilink-tech/omnisim is an open-source robotics simulator built so that coding agents can drive robots and check the results. The latest release is v9.2.6, and the provided materials do not list the changes for that version. Simulations can be controlled and their state inspected over HTTP/JSON and MCP. It is a Python project released under the Apache-2.0 license.

It uses the Newton physics engine for rigid bodies and contacts, cloth, soft bodies, and CUDA particle media. It ships 52 demo worlds, a procedural world generator, URDF and STEP-based CAD import, and robot packages from more than 20 vendors. The ROS 2 sidecar supports the simulation_interfaces standard, and the built-in MCP server exposes 46 tools to a running simulation. On the CPU solver mujoco, it claims bitwise identical reproduction of a contact-heavy 10-robot scene across cold starts. This reproducibility is not guaranteed on GPU solvers. When a sensor cannot confirm a contact, it does not guess the result, and by design the response carries completeness and empty_set_reasons.

In the benchmark, the 30-minute warehouse task _SHIFT had the highest promise-keeping rate at 89.9% out of seven configurations, while the others ranged from 47.9~78.8%. Codex (GPT-6.1 Sol, high reasoning effort), later run with a different model, recorded 76.8% on the same task. That said, the benchmark and the grader were produced by the team itself, so this is not a controlled model comparison. There is no real-world sim-to-real validation, and time control is partial. These numbers therefore show only an in-simulation evaluation and do not demonstrate transfer to real robots.