Meta's humanoid Max, robot orders climb

A social media company quietly built a humanoid, industrial robot orders had their best quarter of the year, and a Korean robot arm maker cut robot debugging time with AI agents. Also: a drone flying through trees without a map, a space station stack on ROS 2 Jazzy, and grenade launchers on ground robots.

Meta built a humanoid called Max, reportedly in three months

The Information reported on Saturday that Meta built a prototype humanoid with Shanghai-based Fourier, internally named Max, in roughly three months, and has started a second version (The Information, TNW summary).

The reported numbers: about 48 kg, around 1.5 metres tall, NVIDIA Jetson Thor compute, five-fingered hands for small objects, a screen face that mimics expressions, and built-in speakers. Max was shown to employees at an internal summit in late August, and Max 2, begun in September, is said to add a new torso, better hands, a projector and sensors for stairs. Meta told The Information the work is internal research with no plans to sell the units, and no launch date has been reported.

Two things keep this in the “reported” bucket. It is a single-source story, Meta and Fourier have published no specs or images, and TNW could not independently verify it. Read it as strategy, not product: a US lab outsourcing a first prototype to a Chinese vendor to compress a build that domestic supply chains quote in quarters.

For developers the interesting part is the compute choice. A 48 kg research platform running Jetson Thor is the embodied version of the VLA and agent pipeline everyone is building, on the same Orin class versus newer edge silicon decision every team faces.

North American robot orders jump 24.8 percent

North American companies ordered 10,989 robots worth $723.5 million in the third quarter of 2026, up 24.8 percent in units and 26.1 percent in value year over year, according to the Association for Advancing Automation (A3).

The mix is the story. Automotive component orders climbed 90.4 percent, semiconductors and electronics 97.6 percent, plastics and rubber 67.7 percent, and life sciences 38.7 percent. Automotive OEM orders returned to growth at 7.5 percent after declining in the first half. Food and consumer goods fell 23.7 percent. Collaborative robots reached 1,831 units and $63.8 million, up about 56 percent and 16.7 percent of all units. Nine month totals: 28,984 robots, roughly $1.89 billion, up 9.6 percent in units and 13.3 percent in value.

One caveat matters for humanoid coverage. A3 counts industrial robots, and a spokesperson confirmed to Manufacturing Dive that the data exclude autonomous mobile robots and humanoids (Manufacturing Dive). So this is the revenue generating base of the industry growing again, driven by chip fabs and battery plants. Cobots at 43.8 percent of life sciences orders is what a shift toward high mix, low volume work looks like in procurement data.

HD Hyundai cuts robot programming time with AI agents

HD Hyundai Robotics says AI agents built with Palantir cut the average time to resolve robot problems from 21.5 days to 7.9 days, about 63 percent, with results tracked since August 2026 (Korea Times, Asia Economy). The company expects to cut programming costs by more than half.

The scope is narrow and that is why it is credible: not robot learning, but the integration work around it. Site engineers spend most of their time on cell configuration, cycle time tuning and fault diagnosis. If an agent can search logs, drawings and past tickets and propose a fix, the 13.6 day saving is plausible without new control theory.

Treat the numbers as company figures with no published methodology or error rate. A useful test would be first time fix rate on faults the agent has never seen. The direction is clear though: the deterministic layer that actually moves the arm stays, and the tooling around it gets a language model.

A drone chases a person through a forest without GPS

Purdue’s IDEAS Lab posted a 44 second video of a quadcopter following a person through trees with no GPS, no map and no cloud link (TechSpot). Professor Aniket Bera says perception and planning run onboard, because a vehicle banking around a branch cannot wait on a server, and that an occlusion prediction step guesses where the target went while they are hidden.

TechSpot adds two details. Purdue announced a five year, $1.5 million agreement with the Army Research Laboratory in late 2025 for machines that navigate without satellites, though the post does not say whether that funded these flights. And the charts on the video label the motion planner “SUPER” and the altitude estimate “FAST-LIO”, both open source projects from the University of Hong Kong’s MaRS Lab, which undercuts the claim that every layer was written from scratch. No paper covering the demo appears on the lab’s publications page yet.

The engineering content is still real: LiDAR inertial odometry, local planning in clutter and target re-identification after occlusion are the hard parts of forest flight, and the same navigation problems show up in ground robots. HKU publishes the SUPER code and a hardware kit.

Space Station OS reaches 0.9.0 on ROS 2 Jazzy

Space Station OS, an open source ROS 2 platform for simulating and operating crewed station subsystems, released v0.9.0 with Ubuntu 24.04 and ROS 2 Jazzy as the native baseline (ROS Discourse).

The release adds a reproducible Pixi and RoboStack environment, so the build, test and station launch commands work without a system wide ROS 2 install. Architecturally it introduces ssos_core for subsystem registration and system management, ssos_sim for world state, simulation time and fault injection, shared station level interfaces, and a Mission Control GUI. ECLSS is the first subsystem ported to the new structure, with EPS, thermal and GNC to follow. The maintainers report 144 tests passing with no failures, errors or skips.

This is worth watching for reasons beyond space. Subsystem lifecycle, heartbeat, telemetry and fault handling, with a clean separation between subsystem logic and ROS specifics, is exactly what lifecycle nodes formalise, under a fault budget where nobody can reboot the lab. Legacy space_station_* packages remain during migration, so expect churn before v1.0.

Ukraine mounts grenade launchers on ground robots

Ukrainian defense companies are putting automatic grenade launchers on uncrewed ground vehicles so operators can fire from tens of kilometers away instead of a few hundred meters (Business Insider). DevDroid’s TW-40 is quoted at 70 km of range on paved roads, up to 50 km off road, with a mounted launcher reaching 1.5 km. The company’s R&D director, Oleg Fedoryshyn, says the appeal is indirect fire: soldiers “don’t need to see the enemy to hit him.”

The technical shift is from teleoperated logistics to weaponized platforms on long radio links, which turns latency, link loss behaviour and fail-safe rules into safety critical design questions. It also shows why import and export policy now treats ground robots as strategic goods.

Quick Bites

• Google plans to support a shared AI training opt-out token in robots.txt, built on the IETF AIPREF work, read as disallowing Google-Extended unless that agent is named explicitly, with support planned for 2027 (Google Search Central).
• A 78 year old retired University of Pennsylvania professor was arrested in Philadelphia on 14 October for standing in front of an Uber Eats delivery robot, and says she will do it again (CNN).
• qerra_gate_node 0.3.0 is an Apache 2.0 ROS 2 Jazzy supervisor that gates /cmd_vel with a lidar stopping envelope and latched fail-closed trips, tested in Webots on a TIAGo base and in CI only (ROS Discourse), a concrete instance of the guard layer pattern.
• Luxonis announced the OAK 4 FFC, a 70 by 53 mm vision computer that runs synchronisation, stereo depth and models onboard, with up to five cameras per robot (IEEE Spectrum).