This week: The Robot Operating System (ROS), the open-source software behind a huge share of the world's robots, spent ROSCon 2026 working out how to make room for AI that learns by watching.

Elsewhere: An open demo that runs a trained robot arm, tidying cubes onto a tray, in a simulator on your own Linux machine.

The New Prime Signal

ROSCon Global 2026, the yearly gathering for developers of ROS, ran September 22 to 24 in Toronto. This week, Hackster.io published a first-timer's tour by Mario Bergeron, who went in with two questions most robot tinkerers are asking right now: are AI models becoming the robot's brain, and is ROS keeping up? The short answer from the talks: ROS isn't being replaced by AI; instead, it's growing the plumbing to carry it. Much of that work comes from a community group that the Open Source Robotics Alliance laid out in August: the Physical AI Special Interest Group. (Also, the ROS mascot is a turtle, and Bergeron opens his write-up with a real snapping turtle hatchling he met near Toronto. It's that kind of conference.)

What shipped: Mostly tools and plans rather than one big launch. The Physical AI SIG works on standard ways to train, simulate, and run AI models on ROS robots, and its members include Intel, NVIDIA, BMW, Intrinsic, Robotec.ai, Ekumen, and FieldAI. At ROSCon, Franco Cipollone of Ekumen and Michael Gentner of BMW Group briefed the crowd on where it's headed. Roboto AI's talk "From Bags to Tensors" covered turning ROS recordings, called bags, into clean training data. Rerun's Isaac Blankenau presented Rosetta, which connects a ROS 2 robot to the LeRobot learning framework through one YAML "contract" that maps robot topics to what the model expects. On the speed side, ROS 2 Lyrical, the latest release, adds native buffer support for passing GPU tensors and image frames between processes without extra copies, and a proposal called REP-0157 goes further on zero-copy messaging. Bergeron also flagged AsyncNode, a talk on native asyncio coming to rclpy, ROS's Python library, which lets a node wait without stalling its other callbacks.

Why it matters: The alliance's post calls ROS "the world's most widely deployed open-source robotics framework," citing its 2025 metrics report. But AI training tools want neat, synchronized frames, and ROS speaks in streams of messages that each arrive whenever their sensor feels like it. Bridging that gap means years of tested drivers and controllers don't get thrown out every time a new model shows up. The workshops had a twist too. In AMD's session, Bergeron writes, the surprise was "code as policies": instead of a big vision-language-action (VLA) model steering the motors, a coding agent writes the robot's control code. The RHO project reports 45.0% success (1,351 of 3,000 trials) on the perturbed LIBERO-PRO benchmark, against 12.83% for the π0.5 VLA and 0% for OpenVLA, per its research paper. That's one benchmark, in simulation, from the method's own authors, so treat it as a promising result rather than a verdict. NVIDIA also told Bergeron that, in its experiments, reinforcement learning still beat VLAs for a precise cable-insertion task.

What's next: Per the SIG's roadmap, the low-cost SO-ARM101 arm demo is "the first of many reference applications." Next up: an industrial-grade example inspired by the AI for Industry Challenge from Intrinsic and Open Robotics, then mobile manipulators, legged robots, and humanoids, plus reinforcement-learning workflows. The zero-copy and scene-description proposals (REP-0157 and REP-0158) are still drafts open for comment, so the plumbing isn't finished. Watch whether robot makers adopt these standards in shipping products, not just demos.

There's a version of the AI robot story where everything old gets tossed for one giant model. ROSCon told a calmer one: keep the tested parts, add adapters, and let the community argue it out in public. It's less flashy, but turtles tend to get where they're going.

The SIG's reference stack is open on GitHub as ros-physical-ai/demos, and it has a shortcut for people who want to see the whole thing work before recording a single demonstration.

How it works: The demos run the SO-ARM101 arm through a full record, train, and deploy loop with LeRobot, connected by Rosetta, in the Gazebo or MuJoCo simulators or on real hardware. You can record demonstrations with a leader arm, an on-screen marker in RViz, or your phone, and there's even an MCP interface so an AI agent can control or debug the robot. The pre-trained demo skips the slow part: 60 pre-recorded demonstrations (80,421 frames at 50 frames per second) and a trained ACT policy with 51.7 million parameters, both on Hugging Face. The task is "place cubes on tray": pick up three cubes of different sizes and set them on a tray.

Try it: On Linux, install Pixi, then run git clone https://github.com/ros-physical-ai/demos, pixi install, pixi run install-deps, and pixi run build. Start pixi run zenoh-router in one terminal and pixi run so-arm-gz in another to bring up the arm in Gazebo, then follow the pre-trained guide to point the policy runner at the checkpoint. The project says you only need an NVIDIA GPU for AI inference or training, not for the simulation itself, and the policy can fall back to your CPU, just more slowly.