github.devotd.
NewRobotics

robosimstudio/robosimstudio — a Real2Sim toolkit for robot manipulation by agents

An open, conversational toolkit that rebuilds real robot manipulation tasks as physics-based simulations and lets agents observe simulator state and refine results.

Repositoryrobosimstudio/robosimstudio ↗

An open, conversational toolkit that rebuilds real robot manipulation tasks as physics-based simulations and lets agents observe simulator state and refine results.

GitHub description: An open interactive toolkit that accelerates human and agentic physics-based Real2Sim of Robot Manipulation Tasks

stars
★ 197
Forks
13
Language
Python
License
Apache-2.0
Read original source ↗View on GitHub ↗

robosimstudio/robosimstudio is an open, conversational toolkit that accelerates the process of moving real robot manipulation tasks into physics-based simulations and repeating them. This Real2Sim work (reconstructing a real environment in simulation) is carried out either by a person repeating the manipulation by hand or by handing the simulation to a coding agent. Human hands-on refinement and agent-driven repetition are bundled into the same tool.

The physics backends sit on top of NVIDIA Newton and Warp, and MuJoCo Warp, and reference external elements such as Isaac Lab, Genesis, SuperDex, and LeRobot, composing them into a single scene. Simulator state reaches the agent as spatial-temporal text and visual feedback, so the agent directly observes the outcome of the scene it changed. The author list of the technical report, arXiv 2610.18835, includes Erwin Coumans and Miles Macklin of the NVIDIA Newton lineage, together with Ken Goldberg. The toolkit is written in Python, released under the Apache-2.0 license, and has 195 stars. The README contains installation guidance and examples, the technical report, and a project page.

That said, the repository has no releases and no tags, and the only thing verifiable is the examples/tests/bench directory layout. The repository was created on October 12, 2026. Performance figures appear only in the technical report, so they cannot be verified from the repository alone. The task counts also disagree. The paper's contribution list presents 18 manipulation tasks, while the README splash image introduces 48. In the evaluation scores, momentum is 3.9 out of 10, and novelty, technical merit, and quality are 7.0, 8.0, and 7.0 respectively. Until the release history and task list are sorted out, it is hard to call this a tool for use in real robot task pipelines.