Spot for ICE, Digit at GXO, ROS 2 agents

A dual-use argument inside Boston Dynamics, two more warehouse sites for Digit, and $600 million for autonomous earthmoving. The developer items are a free course that ends in a trained VLA and a behavior tree debugger.

Boston Dynamics staff fail to stop the ICE Spot sale

More than 300 Boston Dynamics employees signed a petition asking leadership to cancel the sale of Spot robots to U.S. Immigration and Customs Enforcement, and the company declined (WBUR, Ars Technica). The petition reached executives in early September, and Boston Dynamics says nearly half the signatures were anonymous, against a workforce of roughly 1,400.

The hardware path matters. On September 25 the Department of Homeland Security awarded $1.3 million to Radeco, a Connecticut company, to supply four Spot robots modified and marketed as Cerberus. The contract describes them as detecting radiation, toxic gas leaks and explosive threats, plus remote observation and inspection. Boston Dynamics says the buyer is Homeland Security Investigations, not removal operations.

Leadership stood by the deal. In a letter quoted by both outlets, Amanda McMaster, then interim CEO, wrote that customers must follow terms of sale prohibiting use of the product to harm or intimidate, as a weapon, or to enable one. A spokesperson listed the enforcement options: voiding the warranty, refusing software updates, terminating the license, and no further sales.

The terms of sale are the only control layer, and every lever in it is contractual. A Spot in the field still needs updates, so the vendor’s real power is the update channel, not the hardware.

Agility and GXO add two more Digit sites

GXO Logistics will deploy Digit in Statesville, North Carolina and Lebanon, Tennessee, starting in early 2027 on Digit 4 and moving to Digit 5 when it ships (Agility). The work is e-commerce fulfillment, moving totes and loading conveyors.

Behind it: more than 100,000 totes moved at the original Flowery Branch, Georgia site and more than 65,000 hours of real operation across commercial programs. The 2024 agreement there was the first commercial humanoid deployment and the first humanoid robots-as-a-service contract, with Agility Arc handling fleet management.

Two sites is not scale, but this is the first humanoid contract to expand on operational results rather than a pilot extension. The hard part for integrators is the interface to the warehouse execution system and the safety case for working beside people.

Bedrock raises $600M for autonomous excavators

Bedrock Robotics closed a $600 million Series C co-led by CapitalG and Emergence Capital at a $4.1 billion valuation, taking total funding to nearly $1 billion, with NVIDIA among existing backers (The Robot Report).

The product is a retrofit: sensors and software on excavators contractors already own, so the machine does earthwork without a cab operator while crews work around it. Bedrock says deployments have spread from its western base into Minnesota, Florida, Georgia, Missouri and Virginia, on data center, healthcare and heavy civil sites, with Texas contractor Haydon as anchor partner and autonomous trenching next.

The technical bet is that unstructured dirt is now tractable: localize with no lane markings, follow a design surface, and hold a safety envelope around people outside the robot system. That is harder perception than a warehouse aisle, which is why the money follows it.

OMRON buys OnLogic for $675 million

OMRON agreed to acquire 100% of OnLogic for $675 million, keeping the Vermont company as an independent business with its brand and direct sales model (The Robot Report). OnLogic, founded in 2003, builds configurable industrial computers to order and says it has served more than 70,000 customers.

The logic is that inference is moving next to the machine, into cells where a fanless box survives heat, vibration and dust. OMRON has the PLC and motion layer but not a general purpose computing brand, and OnLogic has the hardware but not the automation channel. That combination is worth tracking against the Jetson options that dominate vision workloads.

Xpanner ships an autonomous pile driver

Xpanner launched the X1 Piling R6, an automated pile-driving machine for utility-scale solar and battery storage sites (The Robot Report). One operator uploads a pile plan to Xpanner Connect, walks the machine to the row, then supervises while it moves between pile positions on high-precision GNSS and drives each post.

Verified specifications: a 97 kW (130 hp) engine, a hydraulic hammer with up to 1,260 J of impact energy and three energy settings, a guide rail that slides 300 mm fore and aft and 500 mm laterally, a 180 degree rotating upper body, piles up to 7.0 m long and a maximum driving depth of 4.8 m. Each pile’s position, depth and plumb is logged to a live project map, and the business model is capex or rental plus a monthly software subscription that follows the machine between projects.

This is the least glamorous autonomy here and possibly the most economically direct. The robot needs no understanding of the site, only centimetre localization, a plan, and a hammer that hits the same point every time. Solar buildout is constrained by crews, not demand, so watch whether logged pile data becomes the acceptance record owners require.

A free ROS 2 course that ends in a trained VLA

Anis Koubaa published a self-paced course that starts at ROS 2 nodes, topics, services and actions and ends with a language model driving a simulated greenhouse robot (ROS Discourse).

The pipeline is concrete. You expose ROS 2 skills as tools an LLM can call, collect 49 human demonstrations totalling 4,228 frames for a tomato inspection task, train a SmolVLA model on that data as the pick skill, then evaluate it against simple baselines. Modules 0 to 2 run on 8 GB of RAM with no GPU, and training the VLA needs an NVIDIA GPU with 4 GB of memory, which puts fine-tuning inside reach of a laptop with a discrete card. The workspace runs on Ubuntu 24.04 with Python and is on GitHub, linked from module 0.

The design choice worth copying is the skill server boundary. The model never emits joint torques, it calls typed ROS 2 actions, so safety and failure behaviour stay in code you can test. That is also the honest limit of 49 demos: enough to show the loop, not enough to claim generality, which is the action space question that decides what a policy does with contact and force. The data side is the refinery pattern at miniature scale.

BTick opens BehaviorTree.CPP debugging to the browser

BTick is an open source tool for monitoring, debugging and recording BehaviorTree.CPP applications in ROS 2 (ROS Discourse, code).

It discovers multiple trees and nested subtrees at runtime, supports breakpoints and single stepping, lets you override node status and inject delays, and exposes blackboard variables for editing. Sessions can be recorded and replayed on a timeline. It speaks its own protocol, a Groot2 compatible mode, and the Nav2 logging topic, so a stock Nav2 stack can be visualized without instrumentation, in a browser or desktop app. Most navigation failures are a tick stuck on RUNNING or a recovery branch that never fires, and overrides turn that into a reproducible test.

Quick Bites

• BMW’s Landshut plant is developing humanoid robot software for component production, covering open platforms, demonstration capture with motion capture suits and data gloves, simulation and motion planning, alongside existing pilots in Leipzig and Spartanburg (Assembly).
• A developer implemented the FAST corner detector with ESP32-S3 SIMD instructions and processed VGA at 30 FPS on the 240 MHz part, roughly twice the speed of a scalar version (Hackaday).
• RAIWAY, a wheeled biped with 20 inch wheels, distils simulated contact events into a latent the controller infers from proprioception alone, and climbs outdoor stairs with risers up to 15 cm. Accepted to RA-L (arXiv 2610.33172).