China blocks SoftBank's ABB robotics deal

Beijing stopped a 5.4 billion dollar sale of an industrial robotics division, Agility’s first Nasdaq session gave the humanoid trade a public price signal, Serve Robotics began selling the plumbing it built for its own fleet, and two ROS 2 changes will break builds for anyone who upgrades. The through line is control: over the data a robot generates, and over which layer of the stack a developer actually depends on.

China blocks SoftBank’s purchase of ABB Robotics

China’s Office of the Working Mechanism for Foreign Investment Security Review has prohibited SoftBank Group’s 5.375 billion dollar acquisition of ABB Robotics inside the country, ABB confirmed to The Robot Report. The Financial Times, which first reported the block, says authorities raised concerns about robotics data being transferred overseas and used to train AI models (The Robot Report).

The jurisdictional question is fair, since this is a Japanese company buying a Swiss one, and the answer is manufacturing footprint. ABB builds robots in Shanghai, and about 90 percent of the machines it sells to Chinese customers are made locally. The division reported a 32 percent increase in business in China in the third quarter of 2026.

ABB says it and SoftBank are in close dialogue and working constructively on the transaction, the standard phrasing for a deal that is not dead but has no path yet. China also blocked Meta’s purchase of Manus in April, and the United States has its own file open, with the Section 232 review of robot imports we covered yesterday.

For developers, the consequence is that industrial platforms are being split along data boundaries. If data from a Chinese installation cannot legally reach a model team elsewhere, expect regional stacks and regional model versions.

Agility’s first day sets a humanoid benchmark

Agility Robotics opened its first session at 11.26 dollars, above the 10 dollar reference price in the SPAC deal, traded between 8.40 and 12.15, and closed at 11.49, about 2 percent above its open (The Robot Report). That is a strong day only against the tape: the Nasdaq fell 5.52 percent, Serve Robotics closed down 12.45 percent, Symbotic 2.47 percent and Intuitive Surgical 2.82 percent.

The comparison to keep in view is Unitree, which listed in Shanghai at 150.80 yuan, touched 1,100 yuan on debut and was down almost 53 percent from that high a few weeks later. A first day measures float and enthusiasm, not orders. Agility’s schedule still targets early customer access for Digit 5 in the first half of 2027.

The listing’s real contribution is disclosure. Quarterly numbers on deployments, backlog and losses will give the industry a reference point it has lacked: what a humanoid fleet earns per site. The Robot Report also reports that Agility is exploring wheel based robots, a story it broke in September, and that chief executive Peggy Johnson has floated a smaller bipedal form factor.

Serve Robotics sells the plumbing it built

Serve Robotics launched a business line called Physical AI Services alongside third quarter results, selling three things it built for its own sidewalk fleet: Serve Link, a connectivity layer that ties field robots to AI running in data centres; a data curator that finds edge cases in recorded fleet data; and a catalogue of real world data from its own robots (TechCrunch).

Bedrock Robotics, which builds self driving systems for construction equipment, is named as a customer, and Serve says it is working with T-Mobile to move AI compute closer to robots. Chief executive Ali Kashani calls it the AWS moment: build the whole application first, because that is the only way to learn what to sell, then open the layer underneath.

The claim worth taking seriously is connectivity. Off board inference only helps if the link survives a dead spot, and exception handling needs a session that resumes rather than a request that fails. Anyone who has run a robot on a cellular modem knows how much engineering sits behind the phrase reliably connect.

No latency, throughput or interface details have been published for the three products, so this is a business line and not yet a developer platform. The test is whether Serve Link ships an SDK that code outside Serve’s stack can call.

rclcpp drops topic statistics

ROS 2 removed built in topic statistics from rclcpp, and libstatistics_collector is no longer part of the core, in pull requests merged on 5 October 2026 (ROS Discourse). The removal applies to Rolling now and to Makoa Mata-mata, due 23 May 2027. Humble, Jazzy, Kilted and Lyrical keep the feature.

The C++ API is gone, including NodeOptions::enable_topic_statistics(), SubscriptionOptions::topic_stats_options, rclcpp::TopicStatisticsState and rclcpp::topic_statistics::SubscriptionTopicStatistics. The messages survive in rcl_interfaces, so tools that read statistics_msgs keep building. Replacements already exist: rclcpp emits tracepoints on publish and on subscription callbacks, and ros2_tracing turns those into rates, periods and latencies without touching node code. The community package ros2_pulse reports live topic rates and can publish statistics_msgs on /statistics.

The justification is measurement. The removal cut roughly 2,400 lines, dropped two dependencies and took a one publisher, one subscriber node from 6.32 to 4.91 seconds of compile time on a Ryzen 9900X, about 22 percent. The maintainers skipped a deprecation cycle, arguing that few projects used the feature and the Makoa cycle is early. If you monitor a fleet by subscribing to /statistics, schedule the work now: our primer on topics, services and actions covers what those counters measured.

Robotiq gripper packages move to a new repository

The robotiq_controllers and robotiq_description packages on Humble, Jazzy, Lyrical and Rolling will soon be released from robotiq/robotiq-ros instead of PickNikRobotics/ros2_robotiq_gripper, with PickNik’s agreement (ROS Discourse). The new repository is a fork maintained by Robotiq, and the same release puts robotiq_driver, robotiq_msgs and the rest of that stack on the build farm for the first time. Kilted is unaffected.

Most of the interface is stable: package names, file paths, the robotiq_gripper xacro macro and its arguments, joint and link names, the robotiq_driver/RobotiqGripperHardwareInterface plugin and RobotiqActivationController. Versions go from 0.0.1 to 1.x, and installing robotiq_controllers now pulls in the driver and messages.

Two changes break things quietly. With mock hardware, /joint_states carries only the driven joint, robotiq_85_left_knuckle_joint, and robot_state_publisher derives the fingers from URDF mimic tags, so code reading the other finger joints sees nothing. In Gazebo the plugin becomes gz_ros2_control/GazeboSimSystem instead of ign_ros2_control/IgnitionSystem, and in Isaac topic_based_ros2_control/TopicBasedSystem. At the time of writing rosdistro still lists Jazzy at 0.0.1-3 from PickNik, so the switch is announced rather than landed. Pin your .repos files or snapshot apt before it does, because apt upgrades cannot be undone. PickNik is the company Qualcomm acquired last quarter.

Quick Bites

• Marco Pastorio released ManyForge, a manipulation framework that treats a cell as components and a program as a behaviour tree, run through stock ros2_control on Mock, Isaac Sim and real hardware lanes, with the same architectural bet as Intrinsic’s industrial runtime: a ROS free core with ROS as an adapter (ROS Discourse, repository). Apache 2.0, pre 1.0, one maintainer, and no functional safety.
• A University of Bristol team built a tabletop array of electric motors linked by a shared mechanical backbone and found muscle like coordination emerging without a controller programming it (Robohub, paper).
• NVIDIA published Vera CPU benchmarks for agent workloads, claiming 1.6 times more agents per watt than the fastest x86 part and 1,000 agent sandboxes started in 10.8 seconds against 88.4 seconds on x86 (NVIDIA). Vendor measured, and the robotics link is indirect.
• Ambi Robotics says its sortation systems do not self improve in the field: retraining on fleet data goes through a controlled A/B test before rollout, and the company forward deploys spares and staff to eight new sites before peak season (The Robot Report).