Siemens closes a solver, robots go local

Today: Teradyne invests in an AI infrastructure manufacturer, Siemens closes the open OpenRadioss solver, an FCC ruling pushes robot intelligence on device, a Carnegie Mellon spinout moves to validate generative AI robot safety, and a patent attorney argues the physical AI race will end up in court.

Teradyne invests in Bright Machines to automate AI infrastructure

Teradyne announced a strategic investment in Bright Machines and a collaboration that folds Teradyne robotics and test hardware into Bright Machines’ software-defined manufacturing platform. The target is AI and data center infrastructure, where design cycles are short and rework is costly. Teradyne is the parent of Universal Robots, which recently shipped its seventh-generation cobot arm (The Robot Report).

The integration list is specific: precision robotic assembly, robotic loading and unloading of test equipment, and autonomous movement of materials. Each produces a separate data stream, and the pitch is one production thread linking design, assembly and electrical test. Bright Machines says it has deployed more than 130 microfactories across over 10 countries.

Chief AI officer James Davidson frames the bottleneck: “What limits automation today is not what a robot can physically do but how much engineering it takes to tell it what to do.” That is the integration layer described in the modular robot stack argument, where the value in this deal sits, not in the arms.

Siemens ends the open OpenRadioss solver

Siemens has ended the OpenRadioss project and closed access to it. Radioss is a long-standing finite element solver for crash and impact simulation, and Altair open-sourced a version under the GNU AGPL in 2022. Siemens acquired Altair in 2025, and as of 1 October the project page says only that OpenRadioss is “transitioning to a new phase,” redirecting users to Simcenter Radioss and a research program (Phoronix). Rocky Linux developers have already started a community fork, OpenCourant.

This reaches robotics more than the crash-test framing suggests. Impact and contact simulation is the same physics that makes contact-rich manipulation and safety validation hard to model, and an openly licensed solver lowers the cost of building those models. It is also a reminder that the tooling layer under what Isaac Sim does is controlled by a handful of vendors whose licensing can change without notice.

The FCC ruling pushes robot intelligence on device

In July the Federal Communications Commission added foreign-produced advanced robotic devices, including humanoids and quadrupeds, to the list of technologies it deems an unacceptable national security or safety risk. That blocks new devices from the FCC authorization generally required to import, market or sell them in the United States, over the concern that foreign-made robots could conduct surveillance, collect sensitive data or be interfered with remotely (The Robot Report).

The pressure lands higher in the stack than core control, where motion and safety monitoring are already local. Interpreting instructions, analyzing surroundings, responding to anomalies and making task-level decisions are the workloads that could move on device, since a robot continuously shipping facility data to an external cloud now forces buyers to account for where that data goes. The argument follows that small language models, fine-tuned on a facility’s data and run on edge hardware, fit better than a general cloud model, the same local split that shapes which Jetson to put on the robot. The Robot Report hosts a webinar on the ruling on 27 October.

A CMU spinout validates generative AI robot safety in simulation

Safeworld came out of stealth with a seed round of more than 12 million dollars, led by Shine Capital and a16z Speedrun, with Box Group and the Carnegie Mellon University Endowment also participating. It was founded by Ding Zhao, who directs the Safe AI lab at Carnegie Mellon, with Kyle Wong and Simo Rachidi. The problem is the one generative AI control introduces: a probabilistic model is not predictable the way a traditional algorithm is (TechCrunch).

Safeworld’s method is simulation with realistic human models. It builds a digital version of a specific site condition, such as a blind factory corner, in a simulator like Genesis or MuJoCo, inserts the robot driven by its real software, and runs thousands of scenarios where simulated people encounter it, including a person tripping and falling. The bet is that robot makers want a third party to validate a safety case, the layered assurance already argued for Digit 5 cooperative safety.

A ROS 2 memory runtime enters public alpha

A developer posted Robot Brain (Monolithos) to ROS Discourse: a runtime that keeps task experience (an episode, where it failed, what a human did, what followed) with the source of every fact, then returns what is relevant when a later task asks under its current conditions. With a model key it drafts candidate lessons for review, and new outcomes revise them. Recall alone proves nothing, the author argues, so a feedback call records whether the planner used what came back.

The interfaces are a local HTTP API on loopback (register a robot, batch events capped at 50 events or 1 MiB, recall source facts, recall reviewed memory, submit feedback, revoke an event), a Python SDK, and a ROS 2 reference bridge with task-level DDS services and a persistent request queue, exercised on Ubuntu 24.04, Jazzy and ARM64. Perception, control, policy and emergency stop stay with the integrator, there is no weight training, and model-assisted steps call the provider with your own Gemini key.

The core is not open source: it ships as a compiled package under an evaluation license, and the repository holds only documentation, examples and the issue tracker, with four commits and no stars (ROS Discourse, repository). The design is the part worth reading, particularly the split between source-bound facts and reviewed task memory, a service-oriented take on the topics, services and actions model.

The physical AI race will be won in the patent office

An opinion piece from Dominic Davies, chief executive of Lightbringer and a patent attorney who has filed more than 300 patents in industrial automation, argues physical AI is less likely to be won by whoever builds the best robot than by whoever owns the technology that determines its behavior. His example is Boston Dynamics’ 2022 suit against Ghost Robotics over core technology governing how a quadruped recovered from a fall, which the two resolved in 2025 (The Robot Report).

The points are startup rather than engineering points. Not every component should be patented, and some inventions should be kept secret, because a patent is a publication. Swarm robotics raises a question the law has not settled: whether a claim can cover the behavior of a swarm or only the machines producing it, with the European Patent Office indicating claims may need to define individual entities, the overall system, or both.

Quick Bites

• ROS 2 Jazzy Jalisco picked up 190 new packages and 312 updates in the 5 October sync, tagged jazzy/2026-10-05, with additions including mrpt_path_planning and roboplan (Open Robotics Discourse).
• Boston Dynamics published the full technical explainer behind the new Atlas hand: 13 directly actuated degrees of freedom, the decision to drop the pinky, backdrivability for impact tolerance, and dense pressure sensors on the fingertips and palm, the primary source behind the hand we covered on 2 October (Boston Dynamics).
• Researchers at the University of Glasgow used augmented reality to overlay prototype faces and voices on the body of a Qoobo cushion robot, testing combinations with volunteers to find which readings of emotion make a zoomorphic robot feel like a creature (Robohub).
• Hackaday flags a new class of tiny decision-making models, including Jev, Kev, Nimble and Laya, that accept text but return only floats and confidence scores for yes or no questions and classification, small enough to run locally, which is the shape of model the local inference argument above needs (Hackaday).