Tactile skins, data factories, welding AI

Three announcements this week described one supply chain from three ends: the skin that measures contact, the factory that manufactures training data, and the cobot that spends both on a weld. The most reusable of the three is also the cheapest.

OpenTac publishes a tactile skin you can order

A team at MIT released OpenTac, a full stack for flexible tactile skins that sense normal and shear force, with the design tool, fabrication files, readout electronics and a MuJoCo simulator all open (project page, arXiv 2610.31661). Code is on GitHub, and the team is showing it at CoRL 2026.

Each skin is a flexible PCB with interdigitated electrodes, a piezoresistive film and a 3D printed lid carrying one soft bump over every 2 by 2 group of taxels. A normal load presses the bump down evenly, so the sum of four readings gives normal force; a tangential load tilts it, and the resistance differential decodes the two shear components. Against an OptoForce reference, a 4 by 4 patch reaches in-sample R squared of 0.95, 0.93 and 0.96 across the three axes, over roughly 0.1 N to 10 N at default gains.

The part that changes a schedule is OpenTac-CAD. Give it a 2D outline, a DXF, an STL or a sentence such as a size 9 shoe insole, and it lays out the taxel grid, routes traces and emits Gerbers, a TPU bump case and a manifest. It covered 75 to 93 percent of the target area on five robot surfaces in under three seconds, and all 31 sensors the team shows were designed this way, fabricated through a board house and assembled by hand. A 60 pin readout board handles a 30 by 30 grid over USB, and 30 pin boards daisy chain over I2C so a hand shares one port.

Sampling matters for transients. Raster scanning a 1024 taxel skin yields 67 frames per second, slow enough to miss impacts. OpenTac-Xampling reads a sparse search sample each frame and grows around contact, reaching 1286 partial frames per second at 16 taxels per frame, and it caught every three millisecond impact in the team’s tests where raster scanning missed most of them.

Shear is the channel cheap arrays usually skip, and the one that tells a gripper whether a part is seated or slipping. The caveat is fabrication and calibration, so this is a bench project before it is a product (robot data background).

Generative Bionics plans a data factory, not a demo

Generative Bionics, based in Genoa, announced a Physical AI Data Factory: a 5,000 square metre hub in Italy that it calls one of Europe’s largest physical AI training facilities for humanoid robots, targeting operations in the fourth quarter of 2027 (The Robot Report). Two more sites are planned, one elsewhere in Europe and one for global markets, with AMD Helios rack scale systems providing compute.

The design is a loop rather than a lab. Partner environments are reproduced physically, a fleet of GENE.01 humanoids works continuously, and each environment is also rebuilt as a digital twin where the same task repeats across thousands of layout, object, material and failure variations. Models trained in simulation return to the robots for validation, and fleet data feeds the next cycle. GENE.01 carries a full body multimodal skin that senses touch and proximity, so recorded signals include position, velocity, acceleration, temperature and contact forces, not just video.

Chief executive Daniele Pucci frames the work as sovereignty over three layers: manufacturing data, manufacturing robots, manufacturing models. He also argues that many Chinese humanoids on display are not certified for industrial settings, which is his company’s position rather than a measured result.

The open question is access. A data factory only helps the wider field if its schema, coverage and licence terms are published, and none of that exists yet. Watch whether partner datasets ship with action labels and force channels, or stay inside the building (the data bottleneck, world action models).

ESAB ships a welding cobot that plans its own path

ESAB introduced ROBBI AI, a cobot for MIG and TIG welding that generates the weld path itself instead of following a taught one (ASSEMBLY Magazine). A camera and laser sensing assess the joint, and the system adjusts travel motion, oscillation and process parameters while the arc runs.

The mechanical package is deliberately ordinary: a Universal Robots UR10e or UR20e arm, an optional seventh axis, a wheeled cart, and a choice of EWM Titan XQ, EWM Tetrix XQ or ESAB Aristo Edge 500R power sources. The intelligence is in the constraint. The operator enters the welding procedure specification, including consumables, material and weld type, and the AI must stay inside it, with ESAB’s InduSuite software policing the boundary. If conditions could compromise the weld, it stops and leaves a recording for review. It learns from completed welds after a simulation based learning period, and runs on a computer in the customer’s facility rather than in the cloud.

That is the architecture industrial robotics has been converging on all year: the model proposes, and a gate holding a written specification disposes (gate article, guard layer). Welding suits it because a defect is expensive. Watch for independent weld quality data, since these are ESAB’s own claims.

Progress probes watch a VLA from inside its activations

Monitoring a vision-language-action policy usually means a second model watching the first, which doubles inference cost. A new paper asks whether the policy already contains what a monitor needs (arXiv 2610.31613).

Progress Probes trains linear probes on activations recorded during demonstrations to estimate task progress, then flags an execution that drifts from the expected trajectory. Tested with the open source pi0.5 policy on simulated and real manipulation tasks, the probes matched or beat vision-language-model monitors at detecting failures, and also beat methods that require failure data. A linear probe is cheap enough to run every step.

The limits are the usual ones: the probe is trained per policy and task distribution, false positive rates under distribution shift are not the headline metric, and no code is linked from the abstract. It is still a cheaper starting point than standing up a second foundation model as a supervisor.

One sentence in, one Gazebo asset out

A community toolchain called gazebo-pack compresses the Blender to Gazebo pipeline into a manifest contract and two commands (Open Robotics Discourse, GitHub). A brief such as a bus at 10.5 by 2.5 by 3.05 metres produces a watertight scene, embedded textures, Python build scripts and a manifest.yaml holding dimensions, mass, collision type and layout. Validation covers watertightness, signed volume, a dimension cross check against the mesh bounding box and origin strategy, and the SDF loads offline in gz sim.

The author’s claims are unverified: a 12 tonne bus on box collision stays stable for 45 seconds of headless simulation, and a Tesla Model 3 on four cylinder tyres rests at the theoretical height. Treat the numbers as his, and the pattern as the useful part. Asset generation is where simulation time disappears, and a manifest a validator can reject beats a folder of meshes.

Quick Bites

• NVIDIA published a walkthrough for migrating a ROS 2 node to the CUDA buffer backend for Lyrical’s rosidl::Buffer, using an agent skill to audit allocations and confirm no payload sized host transfers remain, with a Depth Anything 3 TensorRT node as the example (NVIDIA Technical Blog).
• A browser demo trains a 6.5 million parameter model on generated data to follow pick and move instructions, reports about 95 percent success in unseen scenes, and includes a 3D layer viewer that also renders SmolVLA and Cosmos models (Open Robotics Discourse).
• A quadruped solves a giant Rubik’s cube by itself in this week’s Video Friday, alongside the usual compilation of robot clips (IEEE Spectrum).