TL;DR: Nayantra is an open-source framework that lets you command autonomous robot fleets in plain English. An LLM communicates with Open-RMF through the Model Context Protocol (MCP), routing a single natural language prompt all the way down to Nav2 on real hardware or in NVIDIA Isaac Sim. The full pipeline also runs locally on a laptop in stub mode, no GPU, no RMF cluster required.
Motivation: Controlling a robot fleet today typically requires writing dispatcher code, manually constructing task requests, and deeply understanding Open-RMF’s API surface. I wanted to explore whether a large language model, given the right tools, could act as a natural-language interface on top of RMF, letting operators or developers issue commands the way they’d describe them to a colleague.
What Nayantra does:
-
Accepts free-form English commands (e.g. “Send robot A to delivery point 3 and put robot B on standby”)
-
An LLM parses intent and calls Open-RMF APIs via MCP tool calls
-
Tasks are dispatched through Open-RMF and executed by Nav2 on real robots or inside Isaac Sim
-
Stub mode lets you run and test the entire pipeline locally with no infrastructure
Current status: This is an early open-source release. The stub mode works end-to-end and is the easiest way to get started. Real-robot and Isaac Sim integration is functional but still maturing.
I’d love feedback from folks with deeper Open-RMF and Nav2 experience, especially around task API design, edge cases in fleet dispatch, and ideas for extending the MCP tool surface. Issues, PRs, and blunt criticism all welcome.