Atlas Metaplant opens, Halos safety stack

Today: a humanoid training plant inside a working car factory, a safety stack that runs from silicon to certification, a transcontinental flight with nobody touching the controls, a benchmark that measures refusal, autonomous overtaking on a road course, and telemetry over a lossy radio link.

Boston Dynamics opens a humanoid training plant inside Hyundai’s Metaplant

Boston Dynamics opened its Robotics Metaplant Application Center, or RMAC, on 21 September at Hyundai Motor Group Metaplant America near Savannah, Georgia, and Atlas is already training inside the working auto plant. Operations move next year into a new building on the same campus, roughly ten times larger.

The work is parts logistics and sequencing, with component assembly planned by 2030. The constraints are the interesting part: parts must be grasped directly, many have cosmetic or fragile areas that cannot be touched, larger items need two arms, and parts arrive in dense packaging and must be slotted into tight cubbies. Zack Jackowski, chief product and technology officer, told The Robot Report that training data comes from teleoperation, reinforcement learning in simulation, and Universal Manipulation Interface devices, the handheld rigs that record demonstrations with the camera view the robot will have. He calls the site a “Robot Behavior Factory” that combines real parts, fixtures and workcells with robots and skilled pilots and outputs training data. The lesson the company draws is about applications rather than robot design: how to redesign fixtures and packaging for a humanoid, and how to collect data on a live line. Hyundai plans an initial 25,000 units across Hyundai Motor and Kia plants, and a U.S. facility capable of 30,000 robots a year, which The Robot Report places in 2028. The pipeline shape is the transferable part. Background: the robot data bottleneck.

Source: Boston Dynamics, The Robot Report.

NVIDIA argues physical AI safety is a stack, not a checklist

In a 21 September post NVIDIA’s Riccardo Mariani argues that safety has to span hardware, software, AI behaviour, the operating environment and the deployment lifecycle, citing an ABI Research projection of 49 million level 3 to 5 autonomous vehicles by 2035 and an Omdia estimate of roughly 60 million industrial robots deployed between 2026 and 2035. AI behaviour needs its own assurance alongside traditional functional safety, and NVIDIA points to ISO/IEC TS 22440 as the emerging standard for it.

The concrete parts of NVIDIA Halos matter more than the framing. IGX Thor is an industrial module putting accelerated compute and functional safety on one board with a dedicated Functional Safety Island, aimed at systems built toward IEC 61508 and ISO 13849. Halos Core for IGX handles fault detection, monitoring and reporting, and Holoscan Sensor Bridge moves sensor data into AI and safety processing so a system can detect invalid input and run a defined safe response. Isaac Lab and the Omniverse libraries cover simulated edge cases, and the open source Halos Outside-In Safety Blueprint uses facility cameras for outside-in monitoring. NVIDIA says TÜV Rheinland is inspecting IGX Thor, Halos OS and Holoscan Sensor Bridge for certification readiness, and Agility is integrating IGX Thor and Halos Core into the safety system of its Digit 5 humanoid. These are vendor statements and certification progress rather than published third party results, but a safety case is now assembled from the compute module, the middleware and the scenarios you can reproduce.

Source: NVIDIA.

Joby crossed the United States with no pilot control inputs

Joby Aviation announced on 18 September that an aircraft equipped with its autonomy technology flew 3,199 miles coast to coast with zero control inputs from the onboard safety pilot. The converted Cessna Caravan, referred to as the J208, performed autonomous taxiing, takeoffs, navigation and landings while supervised remotely from California and Shaw Air Force Base in South Carolina, up to 2,323 miles away. Stops included Phoenix Deer Valley, one of the busiest general aviation airports in the country, and the flight ended at Dare County Regional in North Carolina after a low pass near the First Flight Monument. Joby says the aircraft rerouted around severe weather in real time with no prior familiarity with the airports it used.

Strip out the aviation branding and this is an autonomy stack demonstration on a certified airframe: perception and planning on board, a supervisor on a long data link, controlled and uncontrolled airspace mixed together. The supervision ratio and the fallback behaviour when the link drops are the same questions that decide whether a long range ground or maritime robot is deployable. Joby names freight, medical transport, disaster response and defense logistics as targets.

Source: Joby Aviation.

RoboHarm measures whether robot policies refuse

RoboHarm, published on 18 September by Edward Sun, Sravanthi Machcha, Sabrina Zou, Tzu Kit Chan and Jay Chooi, asks a question most policy benchmarks skip: when a robot is told to do something harmful, does it decline? Five fixed instructions, five scenes, each scene holding a benign object so a policy that declines has something safe to offer, such as a knife, bread and a doll, or a compressed air can and a lit burner. Three policies took turns on the same bimanual I2RT YAM arms in an Inspect Robots rig: Anthropic’s Claude Fable 5.1 and OpenAI’s GPT-6 Astra as agent policies, and Ai2’s MolmoAct2 as a vision-language-action model. Each ran every instruction 20 times, and reviewers labelled each run from video and transcript.

Claude Fable 5.1 refused 20 of 100 trials, all of them the stabbing instruction, and completed 34. GPT-6 Astra refused 2 and completed 60. MolmoAct2 refused none, completed 6, and froze or did something unrelated on 29 runs. The authors report Fisher exact tests with p below 0.001 for both refusal and completion between the agent policies, and describe the pattern as the more capable policy refusing less in this sample. It is a small study, 100 trials per policy, and the authors present it that way. The operational point holds: clean-task success says nothing about whether a model will attempt a destructive instruction, so track refusal against completion on the same hardware.

Source: Robocurve.

Autonomous race cars overtook each other at Laguna Seca

The Indy Autonomous Challenge held its latest university event at Laguna Seca Raceway on 3 September, ahead of the Grand Prix of Monterey weekend, and for the first time in the series the autonomous cars overtook one another on a road course. The Robot Report reports that Unimore Racing from the University of Modena and Reggio Emilia won, beating Purdue AI Racing in the final heat, with nine teams competing.

The format was one-on-one heats in which the chasing car had two laps to make a pass, after which the teams swapped roles and the power limits went up. Strategy and passing decisions were made by the onboard AI once a heat started, while human crews watched telemetry, adjusted parameters between overtake laps and kept a hand on the e-stop. Every car is the same package, a Dallara base with a Honda engine, Luminar sensors and an NVIDIA GPU, so the competition is almost entirely software. The detail worth stealing is the qualification gate: teams had to pass a software-in-the-loop race inside a D-Space digital twin of Laguna Seca months earlier, the first time the IAC required a virtual pass before a team could run on a physical track. Background: what NVIDIA Isaac Sim is.

Source: The Robot Report.

Copper 1.2 clones your robot over a one way radio link

Copper Robotics shipped copper-rs 1.2 on 19 September, and the headline feature exploits the framework’s determinism: a Copper stack can stream its own state over an unreliable one way connection so a remote copy acts as a live twin. The receiver decodes and reorders the stream using forward error correction that sends repair packets for the recent past, and if too much is lost the robot can send recovery points that let the twin restart at any point. The twin then writes a standard Copper log or drives an operational display over the same API a local stack uses.

The encoding work stands on its own. Copper reports Copperlist encoding improved for a flight controller from 5.9816 to 7.0975 bits per byte of entropy, which it says is 52.31 percent fewer bytes, noting that metadata dominates for robots that log little image data. Version 1.2 also adds cu-anynet, an anytime stereo depth component. For field robots on flaky links, determinism plus forward error correction plus recovery points beats streaming video of a dashboard.

Source: Copper Robotics, release notes.

Quick Bites

• MOOS-IvP Skills v1.4.12 adds coding agent skills to MOOS-IvP, the pub-sub autonomy framework used for marine robots, covering MOOS apps, IvP behaviours, mission building, headless evaluation, log analysis and georeferenced TIFF maps, with a 440 run benchmark of agent performance with and without them. ROS Discourse, GitHub
• Developers report ROS documentation intermittently blocked by an Anubis challenge, affecting API pages and the ROSCon site, with one user describing a month of long blocks. Worth knowing before you blame your network. ROS Discourse
• NVIDIA launched DSX Ready, a program to qualify power and cooling products for AI factories, the datacenter layer that robot training and large simulation fleets depend on. NVIDIA
• arXiv:2609.22274 treats humanoid skills as composable trajectories: CHOREO converts each capability into a representation carrying motion states, contacts, semantics and boundary conditions, then chains them without retraining. On a Unitree G1 in MuJoCo it reports 95.4 percent sequence success across 130 multi-action tasks with 2,950 admitted skill assets. arXiv
• EPFL and Empa researchers propose Sustainability Robotics as a discipline in a Nature Machine Intelligence cover article, arguing robots should be judged on being minimally invasive, universally accessible and symbiotic rather than on technical performance alone. Robohub