Optimus starts, Cosmos opens the pipeline

Tesla put an Optimus line into Fremont and published the compute bill behind it, NVIDIA turned its world model stack into blueprints developers can run, and the ROS 2 task planning baseline jumped to full PDDL coverage. Also: a humanoid pilot at an equipment maker, a red team built on a Cosmos simulation, and 42 generated Gazebo worlds.

Tesla starts Optimus production in Fremont

Tesla reported on 21 October that Optimus production has begun at its Fremont factory on the high volume lines that once built the Model S and Model X (Tesla Q3 2026 update, TechCrunch). The deck calls them “the first production-quality bots that we have manufactured” and says the process suits a commercial product.

Read the capacity table carefully. Optimus gets two rows, California at “Early Ramp” and Texas at “Construction”, neither with a number, unlike the vehicle lines beside it, and a footnote states that installed capacity is not a production rate. On the call Elon Musk called this early production, said the ramp would be slow and bumpy, and claimed 20,000 units per week by the end of 2027.

The cost sits in the same quarter: operating expenses rose 39 percent to $4.78 billion, capital expenditure rose 110 percent to $4.73 billion, and GAAP net income fell 43 percent to $787 million on flat revenue of $28.05 billion.

The useful number for developers is compute. Tesla plans to reach 400 MW of training compute by the end of the year, with Cortex 1 above 105 MW and Cortex 2 above 120 MW, described as supporting “both vehicle and humanoid robot autonomy software”. Musk also said the AI work for open ended tasks is unfinished: hardware lines first, capability later. Watch whether unit counts ever appear.

NVIDIA opens the Cosmos pipeline

NVIDIA expanded Cosmos with data tooling, training recipes and simulation checkpoints (NVIDIA). Cosmos Data Factory ships five cloud native blueprints: driving dataset generation, still image to video of rare or unsafe events, synthetic defect imagery for inspection, and weather or lighting changes that preserve scene structure.

The training side matters more. Cosmos Framework is open and modular on GitHub, with examples on Hugging Face. NVIDIA says developers and agents can adapt models including GR00T, Alpamayo, Qwen and pi0.5 through fine tuning and reinforcement learning, then compress with distillation and quantization before deploying on NIM microservices.

Two deployments are concrete. ANYbotics generated more than 1,500 synthetic anomaly clips to teach inspection robots to spot liquid spills between site visits. A lab called MicroAGI post trained Cosmos 3 on video of humans manipulating objects, then distilled a 32 billion parameter reasoner into an 8 billion parameter model it says lost no accuracy. That ratio matters for anyone shipping on a robot rather than a data center.

Take the efficiency figures as vendor claims: up to 90 percent fewer repeated test drives at Porsche, up to 2x faster inference from the updated NIM services. The pattern is curated real data, targeted synthetic data for the rare event, then a small distilled model at the edge. Background: the robot data refinery.

Laplace red teams physical AI in a simulation

Ann Arbor startup Laplace Intelligence shipped no robot this week, but its use of world models is worth studying (TechCrunch). Founders Howie Sun and Henry Liu built it on Liu’s adversarial testing research, which reached the cover of Nature in 2023.

The product is an interactive simulation reconstructed from real map data, post trained from NVIDIA Cosmos, in which other actors react rather than being driven through. Agents design corner cases from news, collision reports and map data instead of waiting for a human to imagine them. Customers are Bot Auto and a major Japanese automaker, with early humanoid warehouse work and a San Jose traffic pilot alongside NVIDIA and T-Mobile.

The lesson: pass or fail in a static replay proves little, and a policy that only meets scripted actors looks better than it is. The open question is fidelity, since nobody outside the company has published how well these reconstructions predict real failures, a critique that applies to every interactive world model. Read it next to who disposes when a model proposes.

AGCO puts Digit on a Kansas line

AGCO and Agility Robotics signed a commercial agreement to pilot Digit at AGCO’s Hesston, Kansas plant from 2027 (Agility, AGCO). Initial work is parts sorting and stacking plus tote placement, coordinated through Agility Arc, the vendor’s cloud fleet platform. AGCO joins Toyota, GXO and Schaeffler on the reference list.

Agility cites more than 65,000 hours of real world operation across its fleet, a company figure, and the announcement carries no unit count, cycle time or certified safety claim. A 2027 start means site survey and workflow capture run through 2026, where most humanoid projects lose time. Agility is also mid merger with Churchill Capital Corp XI, expected to close in the fourth quarter and list as AGLT. Hardware context is in our Digit 5 roundup.

For integrators, the question is what the fleet platform exposes: telemetry, task assignment, incident logs and an API a factory MES can call. A pilot without that layer stays a demo.

PlanSys2 4.0.0 covers PDDL 1.2 through PDDL+

PlanSys2, the PDDL planning and execution framework for ROS 2, released 4.0.0 (Discourse, tag). Coverage now runs from STRIPS and typing through conditional effects, derived predicates, numeric fluents, durative actions, timed initial literals, PDDL3 preferences and constraints, object fluents, and PDDL+ processes and events.

Two defaults change. Kairos is now the default planner, described as covering more of PDDL than other planners, faster. The STN behavior tree builder is now the default executor and preserves concurrency and time constraints, the difference between executing a schedule and executing a list. The maintainers also report hardening for restarts with state recovery, lost messages and hostile input, tested with fuzzing across Fast DDS, Cyclone DDS and Zenoh. One release targets Humble, Jazzy, Kilted and Lyrical, landing at each next sync.

This is the substrate under the LLM planner wave: a model that emits PDDL needs a validator and a scheduler that respects duration and resources. If your stack dispatches goals through a behavior tree, compare timing behavior before switching.

42 Gazebo worlds generated from Python and Blender

A maintainer published 42 Gazebo worlds for navigation, exploration, SLAM and LiDAR work under Apache 2.0 (Discourse, GitHub): mazes, offices, warehouses, forests, real world places, moving obstacle scenes and challenge layouts.

The authoring pipeline is the interesting part. Instead of placing models by hand, the worlds are generated with Python and Blender and converted to SDF, which makes variants cheap and keeps the output usable outside Gazebo. The limits are stated plainly: 2D plan based worlds with simple objects, and 3D textured environments are future work.

That saves a week of hand placement. Pair it with a working navigation stack (how Nav2 works) and move the same map into Isaac Sim when sensors need to look real.

Quick Bites

• ROS Developer Studio is an MIT licensed browser workspace for inspecting a live ROS 2 graph, including sensor views, host metrics and a Bag Studio that reads recordings offline without publishing into the graph (Discourse, GitHub). It targets Humble and Jazzy, is read only by default, and bag reading is verified on a Humble SQLite3 recording while diagnostics and Jazzy acceptance remain open.

• VeriCalib treats target based LiDAR camera calibration as a verifiable estimation problem, reporting a calibration state from a structured covariance, with rotation and translation precision resolved separately (arXiv 2610.28791). The useful output is not a lower residual but a direction to collect more data in. No code link yet.

• MIT engineers built a two fin swimmer from a 0.5 mm GelMA film carrying light sensitive muscle cells thinner than a hair, reaching about four body lengths per minute and steering a maze by choosing which fin is lit (Robohub, Advanced Functional Materials, Office of Naval Research funding).