Gorai, Jetson strike drone, humanoid recycling

Six things worth your time today: a robotics framework that treats a robot as a NATS mesh, an autonomous strike demo that ran entirely on a Jetson Orin Nano, the recycling problem nobody has costed for humanoids, two papers on contact and terrain reasoning, and a camera calibration tool that throws away the checkerboard.

Gorai builds a robot as a NATS mesh, not a DDS graph

Gorai is a Go framework that starts from the claim that a single robot is already a distributed system, a network of microcontrollers, single board computers and sensors. Every sensor and actuator becomes a service on a NATS mesh, discovered at runtime and addressed by name instead of wired by hand. Sensors are resources you read, actuators are callable tools, and the project wraps that in what it calls NCP, the NATS Capability Protocol, its own version of the capability model MCP gives AI agents, implemented natively over NATS with no MCP server in the path.

Because capabilities are addressed by name and not by location, one logical robot can span a rover, a drone, a sensor mast and a compute box, and degrade as platforms join and leave. Action logs, state streams and replay are first class, and safety is enforced at the capability node rather than trusted to an agent. The CLI validates a robot description file, runs it without a cross-compile loop, builds for linux/arm64 and introspects a live mesh, the part go build cannot do. The repository sits at 43 stars and 224 commits, so treat it as an early bet rather than a replacement for ROS 2 and DDS in production fleets.

Source: GitHub.

Autonomous target selection ran onboard a Jetson Orin Nano

Ars Technica reported on 17 September that Scaleout Systems, a Swedish company founded in 2018 by Uppsala University researchers, demonstrated AI driven target detection and ranking on a loitering munition under the ALMA programme led by BAE Systems Bofors. The drone detected, identified and geolocated potential threats using onboard computing, ranked an armoured engineering vehicle as the highest value target, flew to it and released an explosive. The mission completed without any communication link.

Scaleout CEO Andreas Hellander told Ars the models are deliberately not frontier models, because they have to fit forward deployed hardware that ranges from small embedded devices to edge workstations. The company’s own February write up, relayed by Tom’s Hardware, is the version developers can actually use: an Airolit S1 airframe ran YOLOv8 Nano on an NVIDIA Jetson Orin Nano at about 30 fps while flying at roughly 20 m/s, with target latency of 30 milliseconds or less, in snow at minus 18 Celsius, and derived range without a depth sensor from a pinhole camera model plus known object size. Recon took about 200 seconds and the mission finished in under 320 seconds.

Scaleout’s FEDAIR project sits in NATO’s DIANA accelerator, reported as 100,000 euros of development funding. The interesting engineering is federated learning: edge nodes train locally and share model updates instead of raw sensor data, and in a June test at a Swedish Air Force base a local node kept running inference and active learning after it lost its link to a central node. The transferable lesson is the constraint set, not the application, and it is the same trade off small Jetson boards impose on any field robot. Tom’s Hardware notes there is no NATO procurement order yet, and the UN expert group on lethal autonomous weapons has recommendations due for review in Geneva in November.

The humanoid nobody has planned to retire

A humanoid contains 10,000 to 15,000 individual components, 200 to 500 of them major sub-assemblies, according to a technical piece in The Robot Report by Robert Belt, who works with the recycling firm Re-Teck. The numbers inside that breakdown are the useful part: 20 to 40 electric motors with harmonic drives, 30 to 50 structural elements held by 1,000 to 3,000 fasteners, 40 to 80 encoders and 50 to 200 sensors, and up to 80 memory and storage devices.

End of life raises four separate problems. Retired units hold proprietary navigation maps, biometric logs and behavioural recordings, so storage has to be destroyed or cryptographically erased rather than resold. Battery packs and any pressurised pneumatic or hydraulic sections retain energy that has to be discharged by specialists. Reclaimed carbon fibre frames carry material fatigue that is hard to certify. And a single humanoid holds 3.5 to 4 kg of neodymium iron boron magnets, more than an electric vehicle skateboard chassis, which bulk shredding cross-contaminates with aluminium, titanium and carbon fibre.

The design conclusion is the one developers can act on: modular cartridges, standardised decoupling joints and no permanent adhesives, so end of life becomes disassembly instead of surgery.

Source: The Robot Report.

PIVOT puts a VLM in the costmap, but only as a fallback

arXiv:2609.20983, submitted on 17 September by Aoran Jiao, Wenda Zhao, Hshmat Sahak and Timothy D. Barfoot, augments geometry based terrain assessment with vision-language semantic reasoning for off-road navigation. The paper’s contribution is not the idea, it is the grounding: the authors measure how well the model’s predicted traversal energy cost, robot vibration and wheel slip correlate with real world measurements, then weight each modality by that correlation into one traversability score. The architecture keeps geometric planning as the nominal mode and only invokes semantic replanning when geometry fails to find a path.

Across five repeated closed loop trials on a mixed terrain route of around 6.4 km, autonomy rose from 59.6% to 97.0%, human interventions dropped from 11 to 3, and mean distance between interventions went from 69.2 m to 412.9 m compared with geometry only. For anyone building on Nav2 or similar stacks, the pattern is worth copying: a semantic layer that is invoked on failure, not a learned model replacing the planner. The same vision-language stack that VLA policies use is doing terrain reasoning here.

Source: arXiv.

ForeTac-VLA predicts contact instead of reacting to it

arXiv:2609.20980, submitted on 17 September by Zhengyu Tao, Xin Li and Xin Wang, argues that reactive tactile feedback is the wrong interface for contact rich manipulation. ForeTac-VLA encodes recent tactile observations into temporal representations, fuses them with vision-language features through bidirectional cross-attention, and adds a transformer that forecasts multi step future tactile states, so action generation is conditioned on both felt and anticipated contact. A ground truth to prediction curriculum stabilises training while the forecasts are still unreliable.

Across four real world contact rich tasks the model reports an average success rate of 95%, 36.25 percentage points above a fine tuned VLA baseline and more than 22 points above state of the art tactile enhanced VLA baselines, with performance holding up under low illumination and visual clutter. Those are the authors’ own numbers across four tasks.

Source: arXiv.

Calibrate a camera with no checkerboard

A community tool posted to Open Robotics Discourse on 20 September removes the printed target entirely. AutoCalib calibrates from a flat, textured surface: one straight on reference shot plus 10 to 20 photos of the same surface, with the camera tilted in both directions except in the reference, since foreshortening is what pins down focal length. You can point it at a ROS bag and it samples frames uniformly across the recording. Pinhole with Brown-Conrady distortion and the double sphere fisheye model are supported, and the output is JSON containing intrinsics and distortion coefficients. There is no camera_info YAML export yet.

The accuracy figures published with the tool were scored on a two board rig that was never used for calibration. On a pinhole camera with all parameters free, the noise floor for the test is 0.538 px, a checkerboard calibration reached 0.563 px, and AutoCalib reached 0.660 px. On the fisheye camera the floor is 0.616 px, checkerboard 0.725 px, AutoCalib 0.770 px, and the authors note Kalibr could not detect their target in a single fisheye image. Those are the tool author’s own measurements, not independent replication, and uploading images to a hosted service is a decision to make deliberately, though the site states uploads are deleted about 10 minutes after processing. Background: what ROS 2 is.

Source: Open Robotics Discourse, AutoCalib.

Quick Bites

• Banner Engineering introduced the S100 Multi-Touch Display, an operator interface with up to five touch zones, haptic confirmation, audible alerts, up to 13 double height characters and 16 RGB LEDs in a 100 mm washdown ready housing. IO-Link and Modbus models let PLC logic drive text, LEDs and alerts, the kind of local machine interface a cobot cell usually still needs. Design World
• The ROS 2 Lyrical sync of 18 September added 46 new packages and 315 updates, tagged lyrical/2026-09-18, including depthai 3.10.0 and depthai-ros 3.4.0 for OAK cameras, a mujoco_3d_lidar package and Analog Devices IIO and IMU drivers. Separately, the open LGDXRobot2 mobile platform (mecanum wheels, RPLIDAR C1, ICM-20948 IMU) announced Lyrical support. ROS Discourse, LGDXRobot2
• MIT engineers built a robotic optics lab that assembles, runs and dismantles an optics experiment, with an arm placing mirrors and lenses held in QR encoded housings so the system knows what it is handling. It is a clean example of perception plus precise placement replacing a repetitive bench workflow. IEEE Spectrum
• Dates for the calendar: Humanoids Summit Seoul on 22 and 23 September, IROS 2026 in Pittsburgh from 27 September to 1 October, and CoRL 2026 in Austin from 9 to 12 November, all listed in IEEE Spectrum’s weekly robotics roundup. IEEE Spectrum