Agility shortfall, Audi humanoids, ATEX quadrupeds

A humanoid maker raised less than promised, a carmaker put wheeled humanoids on parts sequencing, a quadruped became legal in explosive atmospheres, and coding agents wrote simulations that frontier models cannot solve.

Agility raises $465M instead of $620M

Agility Robotics and Churchill Capital Corp XI now expect about $465 million in gross proceeds from their SPAC merger, down from the $620 million projected at a $2.5 billion valuation in June (The Robot Report).

That is roughly $164 million from Churchill XI’s trust account, $201 million from a PIPE at $10 per share led by Foxconn, and $100 million of equity already received. Closing is expected on 4 November, subject to shareholder approval.

The order book matters more. Agility says it holds more than $300 million in committed multi-year orders for Digit 5, with deliveries to early-access customers in the first half of 2027 and general availability at the end of 2027. Digit 5 adds redesigned legs and batteries, and Nyobolt fast charging is meant to push operation past 20 hours a day. The revenue base is small: $1.8 million in net sales in 2025 across nine sites, including Toyota, Mercado Libre and Schaeffler, with GXO Logistics adding sites in North Carolina and Tennessee.

Roadmaps written in 2025 assumed cheap capital and 2026 fleets; this one lands in late 2027. Watch the shareholder vote, then the first Digit 5 deliveries.

Audi puts mobile humanoids on parts sequencing

Workers at Audi’s Ingolstadt plant make roughly 1 million picks a day, sequencing parts so the right component reaches the right workstation in order. Audi is testing two-armed humanoids on wheeled bases from Agile Robots to do that sequencing, with AI software from NVIDIA and simulation from Dassault Systèmes (Assembly).

The work sits in a production lab, not on the line. Before any deployment, Audi is building a virtual twin of the logistics environment in Delmia on the 3DEXPERIENCE platform, using NVIDIA Omniverse libraries, to test feasibility, timings, ergonomics and safety. Agile Robots says it has supplied more than 20,000 humanoids, six-axis robots and mobile robots.

The two hard problems named in the project are familiar ones: semi-rigid parts that deform during handling, and unsorted components in mixed containers. Both are grasp planning under uncertainty, where policies trained on clean demonstrations fail at the edges of the distribution. Validating in a twin first forces the task definition to be explicit, the discipline a data pipeline imposes. Watch whether the pilot leaves the lab, and whether Audi publishes picks per hour or intervention rates.

ANYmal X is certified for explosive atmospheres

ANYbotics launched ANYmal X, which it calls the first legged inspection robot for explosive atmospheres, designed to comply with ATEX in Europe and IECEx internationally (The Robot Report). The company says the program took four years and more than 10 patents.

The constraint is why this is hard: fixed equipment earns an ex-proof rating with heavy enclosures or limited electrical energy. A legged robot cannot use either approach. It needs high-power motors, a large battery, exposed navigation sensors and autonomous docking, inside a frame light enough to climb stairs. ANYbotics says it assessed ignition risk across the robot’s service life with the certification bodies and designed each component for it.

The buyers are oil, gas and chemical plants, where the areas that matter most are off limits to standard robots. ANYbotics cites Siemens for up to $500,000 per hour of unplanned downtime in oil and gas, and NIOSH for a fatality rate among U.S. extraction workers seven times the national average.

Treat the first claim as the company’s; the verifiable part is the certification target and the component-by-component method, a template for taking any mobile robot into a regulated zone.

Embotech has moved 1 million vehicles driverlessly

Embotech and BMW say they have moved 1 million vehicles with automated driving technology across four sites in the UK and Germany, handling finished cars from the end of the line to parking or onto trucks (The Robot Report).

Perception lives in the infrastructure: Hesai lidars are mounted in the production environment, so vehicles need no self-driving hardware. A standardized software interface, offboard computers and servers optimise trajectories 10 times per second, and the system connects to existing drive-by-wire. Embotech holds two TÜV-certified SAE Level 4 systems, Automated Vehicle Marshalling and an Autonomous Tractor Solution, on a shared platform, plus a separate deterministic architecture certified by TÜV SÜD that defines and enforces the safety boundary.

That split, a learned layer proposing and a deterministic layer disposing, is what makes certification tractable, an argument that keeps recurring in robot stack design. Moving perception offboard cuts sensor cost per vehicle and makes calibration a facility task, but it puts network latency inside the safety case. Watch whether the model transfers to ports and distribution centres.

Sereact’s Cortex runs Universal Robots arms

Teradyne Robotics partnered with Sereact to put its Cortex physical AI platform on Universal Robots arms for returns handling, piece picking and kitting (The Robot Report). The companies say hundreds of UR robots in dual-arm setups already run at European sites, including Zalando fulfilment centres in Greven, Germany, and Świebodzin, Poland, where the systems process returned fashion items.

Sereact claims one production model serves every robot with no retraining per item, and that learning transfers across the fleet. That is a vendor statement, not a published benchmark, so the test is a returns season with new packaging and unseen garments.

The commercial point: cobot arms with a general manipulation model, sold through an integrator network, ship now while humanoids get the attention. Watch intervention rates as the Zalando rollout widens.

Coding agents build the simulation

A Carnegie Mellon team released Gym-Anything for Robotics, a pipeline in which coding agents write robot simulations, plus RoboWorld, a benchmark built from them (arXiv, project page).

Selection is grounded in labour data: the authors break 1,016 occupations and 465 everyday activities into 31,961 atomic physical actions, value them by wages and time spent, then pick what to build. They shipped 25 simulations covering manufacturing, construction, food service, laboratories, healthcare and logistics, with physics for liquids, cloth, molten metal, granular soil, pastes and soft tissue, plus 25 test and 175 training RL environments with swappable robots from single arms to humanoids.

The result is a useful negative: the best frontier model, named on the project page as Claude Opus 5.5, solves none of the 25 test tasks and averages 0.06 reward. Pipeline, environments and benchmark are released.

Simulation coverage is a content problem, not an engine problem. Scenes, physics, tasks and rewards are code, so better coding agents mean richer simulations without more artist hours. The caveat is that the same agent loop created and audited the tasks, so difficulty is calibrated to that loop. For teams, turning a real workcell into a test environment moves from scene authoring toward prompt plus review, which shortens the loop before VLA training starts.

Quick Bites

• Percepto launched Axion, a hardware-agnostic inspection platform that ingests data from third-party drones, fixed sensors and existing archives, claiming more than 160 AI models, as the company shifts to software first (The Robot Report).
• Humanoid Dodgeball uses threat-aware reward shaping and adversarial motion priors to teach humanoids to evade projectiles, validated in simulation and on hardware (arXiv).
• socx, a JAX solver for chance-constrained stochastic optimal control, reports up to 98 times speedup over MOSEK and 2,224 times over SCS, and solves horizons to N=1120 where MOSEK runs out of memory at N=320 (arXiv).
• The ROS PMC has a new project lead in @ahcorde, and Robotheca launched as a searchable archive of open source robot hardware and software filterable by bill of materials price (ROS Discourse).