ROSCon Global 2026

Preamble

I was lucky enough to get to attend ROSCon Global 2026 in Toronto thanks to Dexory. The conference took place from the 22nd to the 24th of September 2026, with the first day being workshops, smaller sponsor sessions and casual group discussions before the main two days of talks.

Day 1

This was primarily the day of workshops which I was too late to register for unfortunately but that meant I could pick and choose my way through the Birds of a Feather and Sponsor sessions to ease me into the conference.

Birds of a Feather: Deliberation Community Group: Advancing Task Planning, Behavior Representations, and Skill Execution

This was hosted by Christian Henkel and Sebastian Castro on behalf of the ROS Deliberation Community Group. The session was super casual and split into three sub-groups covering Behaviour Trees, Execution, and AI. In conversation with Christian at the AI table he joined up a connection for me between the work of Daniel Kahneman (Thinking, Fast and Slow is a favourite of mine) and robot deliberation, which I fully intend to explore further.

From Perception to Action: Leveraging Intel’s Heterogeneous Compute Architecture for Physical AI Robotics

Intel were showing off their Panther Lake, Intel Core Ultra Series 3, which was really interesting to see how they partition a typical robotics workload across the different core types available depending on demands, be it inference, real-time control, perception, or the middleware itself. Their manipulation example had their perception (semantic segmentation, online camera calibration, and ESDF scene representation) pinned to their four performance cores, their ROS utilities (camera driver, DDS, and manipulation) pinned to their eight efficiency cores, and finally ROS 2 control running on their four low-power efficiency cores. I think this approach of dedicated cores for the varying types of workload is both important but often an area of optimisation that is typically overlooked in favour of throwing more hardware at a problem.

Canonical’s From 10 to 10,000: Scaling ROS and Ubuntu in Production

This was another effort from Canonical to get the robotics community to use snaps for distribution and they were showing off tools for managing software across heterogeneous fleets. They shared an interesting demo that might be useful to try out if you are at a point of deciding how to manage OTA updates. Unfortunately (no bias at all here) the same thing is also covered by…

Dexory/balena’s Scaling ROS Deployments in Production with Secure OTA Updates Using balena

Sam Duffield from balena ran through many of their long-standing tools as well as some new ones, such as being able to talk to the balena supervisor from ROS.

Sam also mentioned balena’s accelerate program to on-ramp startups to their platform.

ROS-Industrial’s Don’t Go It Alone: Joining ROS Industrial for Fun and Profit

Jerry Towler gave a high energy overview on what the mission of the ROS Industrial Consortium is, what it does and how to join. I had my own motivations for being here as I have wondered about getting some of my PhD work into the ROS-I packages, it turns out that there are three separate Organisations that own code from ROS-I.

TODO what are the and where can you contribute.

Birds of a Feather: Outdoor Localization in ROS 2: GPS, GNSS, and Life Beyond robot_localization

I just caught the end of this and there seemed to be a malaise from people complaining about the hand off between different UTM grids, as well as LiDAR and GNSS based localisation systems. On the second point it does seem to me as though people insist on loose coupling of GNSS within their systems, treating it as a black box that either works or does not, rather than thinking how they might aid their GNSS system with the additional sensor suite or even take over the GNSS solution themselves, given the further contraints that they can place on the solution from their sensors. I’m probably talking a good game here but at some point I need to put my money where my mouth is and work on this…

Day 2

The first day of talks and the start of the conference proper with the exhibitor hall opening, though for me interesting stuff was, as always, in the talks. Live streams from both rooms from this day are available in full, here for room one and here for room two.

Space ROS’s First Flight Mission: Challenge’s & Safety

Dave Coleman and Michael Wrock from Picknik presented a great catch up session for the Space ROS project. I have been personally interested in this for some time as a pre-vetted, reliable subset of ROS for several critical application areas. Space ROS is currently a particular build for ROS 2 Jazzy. It drops FastDDS in favor of CycloneDDS, which is something most of the community has been doing for a long time, though Zennoh is displacing Cyclone for many users. Space ROS also drops ROSbag and Secure ROS too. It is complied against several NASA and JPL static analysis tools (IKOS, Cobra, MISRA) with CI test results all archived to allow for auditing. Critically it uses the same core of packages that are used in main line ROS so the community benefits from any hardening offered by the project.

On top of all this they have a scheduled first flight mission which will use BehaviorTree.CPP, MoveItPro, and ros2_control! Scheduled for late 2027, Fly Foundational Robots will run an inch-worm capable Motiv manipulator and is designed purely on getting flight heritage on several key robot technologies, including ROS. All of which will fit nicely into NASA’s In-Space Servicing, Assembly, and Manufacturing (ISAM) goals.

Accelerated Memory Transports: Hardware accelerated pipelines for physical AI in ROS Lyrical

OK, now for the first of the new toys we get to play with!

rosidl::Buffer backends implement a way for us to effectively zero-copy share large data structures using specific hardware backend implementations, like VRAM. The work has largely been implemented by NVIDIA to allow data to stay on their hardware, rather than having to jump on-and-off CPU at every topic transition. The key piece of machinery here is the abstraction of the hardware vendor buffer backend, allowing mapping of a range of different hardware types to the rosidl::Buffer, which might allow for different robot revisions to leverage similar capabilities from their individual compute components. Checkout the two ROS Discourse posts here and here for more.

On Use Of Redbull: A Nav2 Feature Speedrun

“Steve, where’s the bull?” TODO

The new executor on the block - Is it any good?

I’ve had this one open in my tabs for a few months now. TODO

🤓 Homework 🤓

Later that evening I wanted to start work on a comparison of the two new executors for myself. This work is still on-going but intuitively I feel that the events_cbg will provide much lower CPU usage in most use cases, whereas the callback_isolated executor has more targetted tools to decrease latency at the cost of higher CPU usage. As usual there is no one size fits all solution. Once I finish off this comparison I’ll do a dedicated post, and I should really do one on rosidl::Buffer backends too…

Wrap Up

I’ve come away feeling tired, but invigorated. The pace of development is staggering at times, I think the need for human PR review is the only thing that allows anyone to keep up. Anyways, lots of cool ideas for new posts and investigations that I need to do, and no doubt by the time I have done those there will be twice as much new stuff for me to try and get to grips with. But this is why I love working in this field.

Resources and Further Reading

Hey you!

Found this useful or interesting?

Consider donating to support.

Any question, comments, corrections or suggestions?

Reach out on the social links below through the buttons.