Gazebo armhf EOL, NSF physical AI

A simulation stack drops an architecture, a government funder names robotics as a target, a welding cobot learns to paint a moving part, and VLA security research moves from showing the attack to detecting it.

Gazebo ends 32-bit ARM binary packages

The Gazebo team is stopping production of .deb packages for armhf, 32-bit ARM hard float, and the deadline is Monday 21 September 2026. The announcement on ROS Discourse is explicit that armhf is now end of life across every Gazebo release, including the currently supported Fortress, Harmonic, Ionic and Jetty, the same way an Ubuntu release reaches end of life. Supported Debian and Ubuntu architectures going forward are amd64 and arm64.

The reasoning matters if you maintain hardware for a 32-bit board. Raspberry Pi OS, NVIDIA Jetson and ARM cloud instances all ship 64-bit images today, and new ARM Cortex cores have been 64-bit only since 2023. Official ROS 2 binaries are amd64 and arm64 only, and key Gazebo dependencies such as DART do not support 32-bit targets. armhf builders also consume build farm capacity and delay releases when they fail. Packages already published on packages.osrfoundation.org stay downloadable but will not be updated, and building from source on armhf may still work while being neither supported nor tested.

For robotics developers the practical action is a board audit: if a robot or a CI runner still runs 32-bit userland, it needs a 64-bit image before the next simulation dependency bump. Background on the simulator: what NVIDIA Isaac Sim is and where it sits next to Gazebo in a ROS 2 stack.

NSF opens a physical AI funding topic

The U.S. National Science Foundation announced three additional NSF X-Labs topics on 16 September, completing the topic portfolio for calendar year 2026 as part of the 1.5 billion dollar X-Labs investment the agency started in May. The topic that matters here is Artificial Intelligence for Physical Systems, and the call is open for proposals.

The language is unusually concrete about robotics. NSF is asking X-Labs teams to build early-stage platform technologies for AI in physical systems, naming advanced sensors, robotics, embodied systems, human-robot interfaces and cyber-physical systems, with benefits expected in healthcare, national defense, emergency response, advanced manufacturing and scientific discovery. X-Labs are milestone-based awards for independent teams of researchers, engineers and entrepreneurs, described by assistant director Erwin Gianchandani as high-risk, high-impact work that needs resources beyond conventional funding mechanisms.

Two further topics, From Sequence to Function and Computation at the Limit of Physics, are previewed for later this fall, and NSF is not accepting proposals for those yet. An introductory webinar on the physical AI topic runs 14 October 2026 at 1 p.m. EDT, and responses to the agency’s request for information on future topics are due 30 October 2026. For labs and startups, this is the clearest federal signal yet that embodied AI is a funding category rather than a subfield of general AI.

Hirebotics adds line tracking and a linear rail

Hirebotics showed two extensions to its welding, cutting and painting cobots at IMTS 2026 in Chicago (14 to 19 September), both aimed at work that does not fit inside a fixed envelope.

Line tracking lets the Cobot Painter follow a part on a moving conveyor and hold its programmed paint path while the line keeps running. Direction and conveyor movement are calibrated through the Beacon platform, an encoder and sensor report motion so the controller can compensate as speed changes, and the system detects an arriving part and starts the program itself. The second extension is a linear rail, available in 5 ft (1.5 m) increments or custom lengths, that moves the cobot along a long workpiece instead of moving the workpiece to the cobot. A Hirebotics force and power limited arm reaches 68.9 in (175 cm), which is short of a 12 ft (3.6 m) frame or a 20 ft (6 m) structural assembly.

The rail ships now with manual positioning, position verification in Beacon is planned, and a fully automated version that moves the cobot between stations is planned for the end of 2026. Target industries name the constraint directly: shipyards and marine fabrication, structural and heavy fabrication, trailers and truck bodies, agricultural and construction equipment, tanks and vessels. The developer takeaway is that both features are software problems on top of a commodity arm. Conveyor compensation is a coordinate transform against an encoder stream, and the no-code requirement is why it ships in a cloud programming layer rather than in hand written robot code.

VLA backdoors survive fine tuning, and a defense appears

Security work on vision-language-action models has moved past the demonstration stage. INFUSE, posted in January 2026, attacks VLA base models by analysing parameter sensitivity across fine-tuning regimes and injecting the backdoor into modules that stay largely unchanged, freezing the rest. After the user fine-tunes with clean data, the paper reports mean attack success of 91.0% in simulation and 79.8% on real robot tasks, against 38.8% and 36.6% for BadVLA under the same treatment, while clean task performance stays comparable to a normal model.

In July 2026 a second group proposed the first inference-time countermeasure. TrustVLA studies both BadVLA and INFUSE and finds a recurring internal structure: a compact causal footprint, a small visual support that is attention-seeded and spatially tight, where masking it returns a clean-calibrated evidence score to the normal range. The defense adapts a Dirichlet evidence framework from trusted classification to monitor per-token and per-layer epistemic uncertainty in a vision-language-action policy, detects abnormal evidence evolution, localises the suspicious image region by counterfactual score drop, and repairs the observation by localised inpainting. With a small clean calibration set it reduces attack success while preserving clean task performance, with no retraining, and transfers from OpenVLA and LIBERO evaluation to pi 0.5.

Read alongside the VicOne overview in IEEE Spectrum, the pattern for anyone deploying learned policies is that clean-input evaluation is not a security test. Keep a clean calibration set, monitor policy uncertainty at runtime, and treat pre-trained checkpoints as untrusted inputs. Background: why VLA models matter.

An underwater quadruped with SO(3) attitude control

A paper posted to arXiv on 6 September, Design and Attitude Control of an Underwater Quadruped Robot, describes a reproducible legged robot built to work below the surface, with applications in inspection, environmental monitoring and disaster response. It has been accepted to ISRR 2026 and comes from the Autonomous Robots Lab at the Norwegian University of Science and Technology, the same group behind an earlier LSTM hydrodynamic model for swimming quadrupeds presented at ICRA 2025.

The engineering detail is the useful part for anyone waterproofing a legged platform. The motors sit in custom housings machined from polyoxymethylene plastic, sealed with off the shelf O-rings and dynamic shaft seals rather than custom gaskets. The paper derives a simplified dynamic model that captures how drag on spherical end effectors transmits torque to the floating base, then builds a closed-loop attitude controller using an error formulation defined on the special orthogonal group SO(3), evaluated in simulation and in a water tank where the robot tracks roll, pitch and yaw setpoints.

ROSCon Global 2026 runs next week in Toronto

The Open Robotics weekly update confirms ROSCon Global 2026 in Toronto on 22 to 24 September, with free live streams scheduled to go up from 23 September at 13:00 UTC and continue the following day, so the talks are reachable without travel. Around the main programme there is a Rootly mixer on Tuesday evening, a hackathon and campus bus tour on Friday, an Aerial Robotics Meetup and an Intel and Trossen Robotics mixer on Tuesday night, and NVIDIA meetups in the conference hotel coffee shop.

The regional ROSCons follow quickly: UK in Edinburgh on 21 to 23 October, Singapore on 23 to 26 October, Spain in Valencia on 27 to 28 October, Italy in Bologna on 3 November, Belgium in Nivelles on 25 to 26 November, and Turkey in Istanbul on 3 to 4 December. If you maintain a package, the week after a release is when the issue tracker fills up, so it is worth watching the ROS 2 distro repositories as much as the keynote stream.

Quick Bites

• The ICRA 2026 keynote and plenary recordings are now free to watch, including Ken Goldberg on agentic coding, Barbara Mazzolai on biologically inspired robotics and Roland Siegwart on aerial robots, from the Vienna conference held 1 to 5 June. Robohub
• px4-ros2-drone-nav is a map-free 3D drone navigation simulation with a persistent D* Lite planner and CUDA MPPI, built on ROS 2 Jazzy, PX4 SITL and Gazebo Harmonic, under an MIT licence with 66 GitHub stars. GitHub, ROS Discourse
• Nav2 Ready v0.1.0 is a read-only ROS 2 Jazzy CLI that checks the prerequisites Nav2 expects on a custom robot, covering TF connectivity, odometry, laser scan data, the sensor to base transform, the cmd_vel subscriber, map to odom and lifecycle states, with Apache 2.0 licensing. GitHub, ROS Discourse
• A developer has the standard ROS 2 C client stack (rclc, rcl and rmw_cyclonedds_c) running on an ESP32-S3 over Wi-Fi with Zephyr, joining DDS and RTPS directly with no Micro XRCE-DDS agent, tested in both directions against stock ROS 2 Lyrical. Fixed-size messages and best-effort QoS only for now. ROS Discourse