Robot data, cobot software, local AI

Today: Schneider Electric bids $22.6 billion for PTC, Mecka AI raises $60M for robot training data, Universal Robots ships Gen 7 as a cobot lawsuit ends, and an acoustic sensor catches a slip before the drop.

Schneider Electric to buy PTC for about $22.6 billion

Schneider Electric said this week it will acquire PTC, the Boston-based industrial design, engineering and data management firm, for about $22.6 billion. The two companies describe the combination as a “digital thread” running from CAD to the shop floor, promising faster robot testing in simulation and easier maintenance (The Robot Report).

For robotics developers the interesting part is the tooling. PTC owns Onshape and Creo for CAD, and Codebeamer for requirements, testing and traceability on software-heavy products. Schneider Electric already holds AVEVA for industrial operations. Interact Analysis analyst Matthieu Kulezak told The Robot Report that PTC extends Schneider from factory operations and energy management into product design and lifecycle management, and gives it access to robotics developers and machine-building teams earlier, before those customers choose automation equipment. Siemens and Dassault Systemes lead in simulation, so the deal narrows a portfolio gap rather than closing it.

The execution risk is real: Schneider Electric shares fell sharply on the announcement, and Kulezak points to the 42.3 percent premium, the financing and the integration work. Watch whether Onshape and Codebeamer stay open to third-party robot stacks.

Mecka AI raises $60M for human motion data

Mecka AI, founded in 2024, announced a $60 million Series B led by Sequoia, with participation from NVIDIA and Microsoft’s venture fund M12. TechCrunch had previously reported that the round was being raised at a $500 million valuation. The company pays people to record everyday tasks while wearing body sensors and using smartphones, then turns that footage into training data for humanoid and other robots (TechCrunch).

The bet is that robot data will follow the LLM data market. Mecka positions itself against Scale AI and Micro1, which began in text labelling, and against XDOF, which TechCrunch previously reported was in talks for a Series B at a $1.2 billion valuation. The investor list is the strategic detail: NVIDIA and Microsoft are both building robot models and both want the capture layer underneath them.

For developers, the open question is whether consumer-grade capture yields the state and action pairing that imitation learning needs, or just more video. That is the robot data bottleneck in one sentence: collection is now well funded, but annotation, alignment and provenance still decide whether a dataset actually trains a policy.

Universal Robots ships Gen 7 as a cobot lawsuit ends

Two Universal Robots stories landed in the same week. Teradyne Robotics and Elite Robots said they settled a software dispute that had produced a preliminary injunction from the Regional Court of Hamburg in April. In a joint statement the two companies said the settlement “does not constitute an admission by Elite Robots regarding liability, infringement, validity of intellectual property rights, or any other disputed issue,” and neither side disclosed the terms (The Robot Report). A separate Teradyne case against Chinese cobot maker JAKA Robotics’ German subsidiary over three Universal Robots patents remains open.

Separately, Universal Robots announced Gen 7, a redesign of its force- and power-limited cobot line built around PolyScope X, which the company calls the software platform for its new generation. The platform covers new g-Series arms, a CB7 Core controller, a TP7 Core teach pendant and an SP7 Smart Panel, with open APIs and an SDK aimed at AI applications. Parent company Teradyne says it has sold more than 100,000 collaborative arms since 2005.

The timing is pointed: the dispute concerned Universal Robots software, and the new platform is a software platform. Open APIs on the controller are the part developers should read closely, because they decide whether a cobot arm can host learned skills the way a modular robot stack expects.

HUMXN pays for home repairs to harvest robot data

HUMXN, an AI data company based in Minneapolis, launched a program that pays for home-service appointments so technicians can record them as training data. Qualifying homeowners in Minneapolis, Chicago and Miami receive free plumbing, electrical, HVAC, air-duct cleaning and carpet-cleaning services, while participating technicians wear capture hardware that documents the job from their point of view (HUMXN).

Founder Thomas Unise frames the program as turning skilled labour into a structured dataset. The value sits in the process rather than the finished repair: what a technician examines, which tools they choose, how they respond when conditions change. That is closer to the demonstration data imitation learning wants than an instructional video, and it is the same target that teleoperation and rosbag2 recording workflows chase.

NVIDIA puts petaflop AI on Windows desktops

NVIDIA and Microsoft announced RTX Spark, a laptop and compact desktop platform that brings the NVIDIA AI stack into Windows machines. RTX Spark pairs a Blackwell RTX GPU with up to 6,144 cores and an up to 20-core Grace CPU connected at 600 GB/s, with up to 128GB of unified memory and up to one petaflop of FP4 compute. Laptop preorders opened the same day with shipping on 16 October, and compact desktops follow in November. NVIDIA also previewed DGX Station for Windows, built on a GB300 Grace Blackwell Ultra desktop superchip with 748GB of coherent memory and up to 20 petaFLOPS of FP4, which NVIDIA says is enough for trillion-parameter-scale models locally (NVIDIA).

Microsoft used the same event to declare general availability of Microsoft Execution Containers, OS-level infrastructure for running agents in the background under operating system control.

The robotics bridge is memory and locality. Larger world models and conditioned policies have been pinned to cloud GPUs or Linux workstations, and the same CUDA stack now targets a Windows desk with unified memory. DGX Station for Windows keeps Linux toolchains reachable through WSL, so the practical question is driver, container and SDK support for the ROS 2 and Isaac stacks a robot team already runs.

Acoustic sensing catches slip and fracture before the drop

A new paper presents SAFE, a grasp-state detector that uses two passive polyvinylidene fluoride (PVDF) acoustic sensors plus the gripper’s own motor proprioception, with no camera and no prior knowledge of the object’s material. The sensors mount on a compliant Fin Ray gripper, and a single HistGradientBoosting classifier labels the state as normal, slip or fracture from a 79-dimensional feature vector (arXiv:2610.08802).

The numbers are the story. Under leave-one-grasp-out cross validation, SAFE reports an Alert-F1 of 0.884 with near-zero slip-fracture confusion, and ablations confirm the acoustic channel is doing the work. An adaptive grasp controller built on that detection layer runs at 104 Hz on a Jetson Orin Nano and reports 91.3 percent success across 46 closed-loop robot trials, against 0 percent for fixed-force strategies on object conditions that mismatch their preset, and 82.4 percent on objects never seen in training. The authors are Zerun Wang, Vivek Kamat and Shekhar Bhansali.

Grasping fragile objects without object-specific calibration is the gap this addresses. Two cheap passive sensors plus a small tree-based classifier is a different cost profile from tactile arrays or vision-based slip detection, and 104 Hz on an Orin Nano fits a normal edge controller. Watch for a code release and for replication on grippers other than Fin Ray.

Quick Bites

• A framework using vision-language models to combine crop type, pesticide type and weather data for adaptive spraying reports at least 30 percent better crop-row detection and reduced drift, with a Model Predictive Path Integral controller for navigation; accepted in Computers and Electronics in Agriculture (arXiv:2610.08807).
• An autonomous droplet navigation platform drives liquid across an open, tilting board coated with a thin silicone oil film, learning a model-based RL policy in 50 to 150 physical episodes with no simulator, transferring zero-shot to unseen geometries and completing training in under 90 minutes (arXiv:2610.08852).
• Go2-DrivoR adapts an end-to-end autonomous driving planner to quadruped urban navigation by conditioning trajectory generation on a local-frame subgoal, trained only on TartanGround simulation data and transferred zero-shot to open-loop real-world trajectory prediction (arXiv:2610.08812).