Multi-floor SLAM and a blocked robot deal

An open source SLAM stack treats every floor as its own map, China’s security review stops SoftBank from buying ABB’s China robotics business, and Meituan publishes the data recipe behind a desktop manipulation model. Also: Serve Robotics sells its fleet plumbing, and a Unitree factory tour runs into the company’s own IPO filing.

StrataSLAM keeps floors apart in ROS 2

The Intelligent Robotics Lab at Universidad Rey Juan Carlos released StrataSLAM v0.2.0 on 9 November, an Apache-2.0 multi-floor 3D SLAM stack for ROS 2 Jazzy on Ubuntu 24.04 (announcement, repository).

The core choice is a floor gate: one submap per floor, and no online code path accepts a loop closure between two floors, so identical office levels never merge. Stairs, ramps and elevators are detected and classified. On an elevator ride the cabin moves with the sensor, so the system integrates the IMU to measure vertical motion while the camera or lidar sees only the cabin walls. An offline fuse step then places floors in a gravity-aligned four degree of freedom frame using transition constraints, an optional barometer, structural anchors and a weak floor-height prior, even without visual overlap.

Inputs can be RGB-D, 3D lidar, or both, with IMU, barometer and wheel odometry optional. One YAML profile per robot covers ground, handheld and aerial templates, and strataslam doctor checks topics, rates, TF and time sync before a run. The node publishes map -> floor_ -> odom following REP-105, so Nav2 and RViz see a normal map frame while mapping, with publish_tf: false when AMCL already owns it. Maps export as one PCD per floor, or PLY, and the repository ships a simulator with fifteen scenarios plus runs on real elevator data and ramps.

Multi-floor operation is where many stacks quietly fail. If the floor gate holds in your building, it removes a custom layer between SLAM and the Nav2 stack. At v0.2.0, expect interface churn.

China blocks SoftBank’s ABB Robotics purchase

ABB Robotics confirmed to The Robot Report that China’s Office of the Working Mechanism for Foreign Investment Security Review blocked SoftBank Group’s planned acquisition of ABB Robotics, the division ABB agreed to sell in October 2025 for about 5.3 billion dollars (The Robot Report).

The reported concern is data: whether robotics data and assets developed in China could move overseas and train AI models, with SoftBank’s stake in OpenAI part of the backdrop. ABB and SoftBank said they remain in close dialogue, and neither has published a detailed regulatory finding. The prohibition covers the China business; the companies have not announced that the global sale is cancelled. ABB Robotics has installed more than 500,000 robots and employs about 7,000 people.

The context matters. The Robot Report notes that the US Federal Communications Commission recently banned new models of certain connected devices, including legged robots and AMRs, and that the Department of Commerce determined imports of robotics and industrial machinery threaten the US economy and national security. Industrial robotics is being treated as strategic infrastructure at the intersection of automation, data and AI.

For teams, merger clearance is not the only gate: a deal involving local factories, engineers and operational data can be stopped after competition review.

Meituan trains a desktop VLA on 15,060 hours

Meituan’s robotics team published Meituan-Robotics-0, a vision-language-action model for stationary-base desktop manipulation, with code and weights linked from the project page (arXiv, project page, code).

The numbers matter. The corpus combines 147 million vision-language samples with 15,060 hours of tabletop trajectories across 12 embodiments: 7,530 hours of licensed real-robot data, 3,235 from RoboDesk and 4,295 from simulation, leaving 10,765 hours of physical execution. Segments with base motion are excluded, keeping the policy scoped to a fixed base.

Training runs in three stages: discrete FAST action-token pre-training, flow-matching mid-training on a unified end-effector representation for cross-embodiment control, and closed-loop post-training on real robots. A method called Rollback keeps the policy segments that correction replaced and uses them as negative supervision for CFG-DAgger, so the model learns from the failure as well as the fix.

The authors report 93.6 percent success on RoboTwin 2.0, an average of 14.63 percent on RoboDojo (fifth in its VLA track), and better average success and progress than pi-0.5 under matched HG-DAgger post-training on three real tasks. Ethernet insertion reaches 70 percent success with a progress score of 0.90. These are the authors’ measurements, not an independent reproduction.

The reusable parts are the unified end-effector interface that lets one policy span 12 robots, and the event-driven agent that keeps persistent task events so a long job resumes after a temporary blockage, both covered in our robot data refinery piece and the action space and force discussion.

Serve Robotics sells its fleet plumbing

Serve Robotics launched Physical AI Services alongside third-quarter results on 5 November, selling the connectivity and data tooling it built for its own sidewalk fleet (TechCrunch).

Serve Link connects field robots to AI systems in data centres, aimed at unreliable wireless links and exception handling. A data curator finds edge cases in footage a fleet already collects, and a third product is a real-world data catalog built from Serve’s sidewalk robots. Bedrock Robotics, which builds self-driving systems for construction equipment, is a customer, and Serve is working with T-Mobile to move AI compute closer to robots.

Chief executive Ali Kashani calls it an AWS moment: build the end-to-end application first, then open the platform once others want to build similar things. The question is whether a mid-size robot maker wants a competitor as its infrastructure provider.

Unitree’s factory tour versus its own filing

Josef Chen toured Unitree’s Hangzhou factory in early November and published an account that contrasts the company’s automation marketing with its IPO prospectus (write-up).

In the actuator building he describes one or two new automated lines, with the rest of the floor staffed by people running conventional copper winding machines. Unitree has promoted a video of a humanoid assembling robots in its own factory as an actual deployment. Chen reports that the prospectus calls it a pilot and states that assembly is currently mainly manual. He also reports that the book value of all Unitree machinery at the end of 2025 was about 1.7 million dollars, against roughly 32,000 robots built that year.

The economics explain why. Chen reports labour at 8 percent of unit cost, with parts making up most of the rest, so halving assembly labour saves about 4 percent. Unitree expects its new motor line to cut labour per motor by more than half and the cost of one robot model by under 2 percent. The filing also shows where demand comes from: three quarters of humanoid revenue in the first nine months of 2025 came from research and education, under 5 percent from factories, and half or more of what is filed as industry is guided tours.

This is one engineer’s reading of a prospectus, not an audited breakdown. The point stands: humanoid cost is a bill of materials problem with a software problem on top, not mainly an assembly automation problem.

Quick Bites

North American companies ordered 10,989 robots worth 723.5 million dollars in Q3 2026, up 24.8 percent in units year over year, with cobot orders up about 56 percent to 1,831 units and automotive component suppliers up 90.4 percent (ASSEMBLY).

NVIDIA and LG Electronics ran an Isaac GR00T training loop for the CLOiD robot in a real manufacturing environment, chaining real data, Cosmos augmentation, GR00T training, Jetson AGX Thor inference and simulation-based evaluation, while Kakao Mobility builds level 4 driving on DRIVE AGX Thor (NVIDIA).

Physicl raised a 10 million dollar seed round led by Inovia Capital and released Studio, a library of 10,000 3D environments and 100,000 assets for synthetic robot training data (TechCrunch).

ForPC, a force-aware whole-body controller for humanoid cart pushing, cut cart-velocity RMSE by up to 68.2 percent against a locomotion-only baseline and transferred from simulation to a Unitree G1 without real-world fine-tuning, pushing 30 kilograms at 0.5 metres per second (arXiv).