Ex-proof quadrupeds, batched planners, agent harnesses

Two industrial deployments that took years to certify, plus three developer releases that lower the cost of planning and wiring a language model to a real arm. In all five, the engineering is in the constraints.

ANYbotics certifies a quadruped for explosive atmospheres

ANYbotics launched ANYmal X, which the Zurich company calls the first legged inspection robot certified for explosive atmospheres, and says deliveries to customers have started (The Robot Report).

The certification is the story. Equipment in a hazardous zone must be explosion protected under ATEX in Europe and IECEx internationally, which fixed equipment achieves with heavy flameproof enclosures or by limiting electrical energy. A walking robot can do neither: it needs high power motors, a large battery, and exposed cameras for navigation, all in a frame light enough to climb stairs. ANYbotics says the robot computer and inspection payload sit in nitrogen filled sealed chambers, non-metallic shells insulate metal parts, the feet are conductive for grounding, and all 12 leg actuators and the battery are inside flameproof casings. Charging and data transfer use energy barriers so no connector is exposed to the atmosphere.

The company claims four years of engineering and more than 10 patents on the design. It says more than 50 customers run over 120,000 autonomous inspections a month on its Shift platform, which connects findings to SAP, Yokogawa and SLB maintenance systems. BP, Shell, Equinor, BASF and SLB tested it through an early access program since 2023.

Certification shapes the bill of materials before any perception code is written. The inspection data, not the locomotion, is the likely moat.

BMW moves a million cars with offboard autonomy

Embotech and BMW say they have moved 1 million vehicles with Embotech’s Automated Driving In-Plant system across four sites: Dingolfing, Leipzig, Regensburg and MINI Plant Oxford (The Robot Report).

The architecture inverts the usual self-driving stack. Perception lives in the factory, not the car: Hesai lidars are mounted in the production environment, offboard computers optimize trajectories 10 times per second, and vehicles need only drive-by-wire, with no added sensors. A separate deterministic layer, certified by TÜV SÜD, defines and enforces the safety boundary around a SAE Level 4 operation.

The numbers are operational: about 3,500 vehicle movements a day at more than 99.5 percent availability, and the second 500,000 movements took 7 months against 15 for the first. Embotech says the next site is BMW Spartanburg.

This is the cheapest autonomy to certify, because the perception hardware is fixed and maintained by the plant. It is also the least portable: intelligence in the infrastructure keeps the robot there.

ROBOTIS hands a general model a ROS 2 tool belt

ROBOTIS published a technical story and code for running GPT-6 Astra on its AI Worker humanoid through the Codex App Server, with no task specific policy training (ROS Discourse, docs).

The harness is the useful part. The model picks a tool and its inputs; the App Server routes the call to a harness that validates them and the execution conditions, then publishes a ROS 2 message. Arms take a robotis_interfaces/msg/MoveL goal on /l_goal_move and /r_goal_move, grippers and lift take joint trajectories, and the base takes a geometry_msgs/msg/Twist on /cmd_vel. After each command the harness checks the actual robot state, collects images from three cameras plus robot state, and returns the result for the next decision.

Guard rails are explicit configuration: limits.max_step_m is 0.1 m per move, a plan holds up to 5 steps (plan.max_steps), and episodes are bounded by episode.max_actions and episode.max_seconds. ROBOTIS reports that shrinking max_step_m cost tokens and episode time but improved success, apparently because smaller moves create more chances to look and correct. The documented limits are honest: depth from 2D images, collisions, repeatability, latency, and the fact that Ctrl+C stops new actions without stopping the robot.

That is the propose and gate pattern end to end: the model never emits joint torques, and every command passes a validator you can test. The same structure runs in Gazebo or on the physical SG2, and the code is on GitHub.

batchplan batches grasp planning inside MoveIt

A new open source project, batchplan, added a ROS 2 entry point that takes several candidate grasp poses in one request, solves IK for all of them, plans to each in a batch, and returns the shortest valid path (ROS Discourse, code).

batchplan_planner speaks MoveIt’s MoveGroup action and GetPositionIK service without needing MoveIt, and an optional pipeline plugin swaps OMPL for batchplan in RViz. The core is a Rust library with optional wgpu acceleration for IK and trajectory optimization on Vulkan, Metal or DX12, or plain CPU threads with no CUDA. ROS 2 Lyrical Luth and Jazzy are both supported.

The published timings, on an Apple M4 Pro CPU, change the calculus: 64 grasps planned in one request take 18 ms against 620 ms when sent one at a time, and 8 grasps take 8 ms against 62 ms. Through MoveIt’s /plan_kinematic_path, 20 alternating requests around a wall were solved 20 of 20 at 3.8 ms median planning time, against 15 of 20 and 16.4 ms for OMPL’s RRTConnect, with shorter paths, 2.69 rad against 3.72 rad. It keeps the jerk limits, so free space motion is slower.

Early: sphere based collision geometry and an rclrs 0.8.0 workaround on Lyrical. Still, best of N grasp selection is where sampling planners waste time, and batching it is cheap on hardware most labs own. It lands next to the MoveIt ownership change as evidence the planning layer is being rewritten from below.

cuArms plans eight arms with GPU conflict search

Researchers released cuArms, a conflict based planner for multi-arm workcells that is co-designed with GPU parallelism and paired with a task scheduler (arXiv 2610.36250, project page).

The method borrows from multi-agent path finding: each arm gets a reusable multi-modal roadmap, conflicts and deadlocks are resolved with a LaCAM style lazy search over GPU batched conflict tables, and trajectories are shortcut afterwards. A scheduler proposes task skeletons, which cuArms grounds into coordinated plans with handovers and carries, and replans when a skeleton proves infeasible.

The baseline is RoboBallet, a reinforcement learning method for the same 4 and 8 arm spot welding benchmark with 10 and 40 targets. cuArms solved 150 of 150 episodes in every setting tested, where the retrained RoboBallet policies solved 29 of 150 for 10 targets and 50 of 150 for 40, and reached a first solution in 0.20 to 0.87 seconds median on an RTX Pro 6000 workstation, against roughly two days of retraining. On the 10 target problem the initial makespan of 10.61 s improved to 5.64 s.

The authors also run it on two Flexiv Rizon4 arms doing electronics assembly. If the code lands, the use is feasibility checking: whether a cell layout and arm count can hit a cycle time before committing to hardware, a different question from what action space choices answer.

Quick Bites

• FAW-Volkswagen is expanding its Ubtech Walker S Lite trial from vehicle inspection into logistics, after a cooperation agreement with Ubtech this month; no robot count, factory or timeline was disclosed (Assembly).
• E0.5 puts a 3D foundation model in the backbone of a world-action model and predicts future geometry alongside actions: 89.90 and 50.09 percent on RoboTwin 2.0 clean and randomized, and about 15 and 7 point real robot gains after 160k AgiBotWorld-Beta trajectories (arXiv). Background: the world-action survey.
• A Gazebo Harmonic plugin models a marine radar for a USV, publishing a PPI image, a radar return point cloud and a target list, with configurable range (ROS Discourse).
• A ROS 2 Jazzy C++ driver for the Blueprint Subsea Oculus M3000d imaging sonar is open source, with a known bug clamping ping rate reconfiguration to 9 Hz (ROS Discourse).