Open 7B action model, all-linear hand

Today: an open 7B model takes the top spot on a public manipulation benchmark, a Duke lab argues dexterity does not need 20 motors, and two Y Combinator robotics companies ship developer tooling.

Black Forest Labs opens a 7B world action model that tops RoboLab

Black Forest Labs published FLUX 3 Action on 23 September with Hugging Face and NVIDIA: an open weights 7B world action model that takes one or more camera frames, a state vector and a text instruction, then returns the next 2 seconds of actions as 32 action tokens plus, optionally, 32 decoded frames. It is a diffusion transformer where a frozen video VAE encodes the frames and a frozen Qwen3-VL-4B encodes the instruction, with actions carried as a second token stream, one token per future frame, denoised jointly with the video tokens.

The team fine-tuned it on DROID and reports first place on the RoboLab-120 benchmark at 42.92% success, ahead of the 36.8% the post lists for an open 16B Cosmos 3 Nano policy. DROID and SO-101 checkpoints ship as a LeRobot policy class with parameter-efficient fine-tuning recipes. In their own clips, a policy adapted on about 200 teleoperated episodes picked objects it had never trained on and kept working when the camera moved between training and rollout.

An open 7B model that beats a 16B policy on a shared benchmark changes the compute math for teams fine-tuning their own. Treat it as the authors’ reported result, and see why vision-language-action models matter for where these policies still break.

The Cartesian Hand gets dexterity from linear motion alone

A Duke University group (arXiv, posted 22 September, Boxi Xia and colleagues at the General Robotics Lab) built a 7 degree of freedom end-effector with no articulated finger joints. Two independently actuated parallel grippers hold different parts of an object while four translating fingertips push those parts against each other, so manipulation is composed from linear motion primitives with configuration-independent kinematics.

That covers the objects that normally demand a second gripper, an external fixture or extra arm motion: threaded containers, pivots, linear guides, plungers and triggers. The paper demonstrates cap opening and closing, pipetting, pumping, screwdriving, two-handle manipulation and in-grasp reorientation across 35 laboratory, manufacturing and household objects, then transfers the same procedures to a humanoid with a hand on each arm. The authors say they will open source the software and the hardware design.

If linear actuators can produce in-hand manipulation, dexterity stops being a motor count problem and becomes a mechanism design problem, which pulls down cost, control complexity and the number of parts that can fail.

A YC startup builds a low-cost humanoid for developers

Nori Robotics, a Y Combinator S26 company, describes its A3 platform for developers and home tasks: two arms with 7 plus 1 degrees of freedom each and 1.5 kg payload per arm, a lidar with 12 m range at 8 to 12 Hz scanning, four 720p RGB cameras up to 30 fps on the grippers, head and neck, a speaker and microphone, and a 6 to 8 hour battery. A laptop app, Nori Lab, handles training and operation, and a skills marketplace is planned so owners can share what their robot learns.

Payload per arm decides which manipulation tasks are possible at all, and the task list on the site, kitchen help, tidying, fetching, folding clothes, pouring ingredients, is a roadmap rather than a measured benchmark. What to watch is whether the software stack is open enough for teams to bring their own policies.

HFlow opens the data plumbing every robotics team rebuilds

Hebbian Robotics, another Y Combinator S26 company, published HFlow, an open source SDK for multimodal robotics data pipelines. Its v1 boundary is MCAP, the container ROS 2 records natively, written with in-band H.264 whose GOP length matches how the data is read, plus topic group chunking so camera and state streams never share a chunk.

Processing stays in your own Python functions, transforms, checks, labels and enrichments, run in-process during development or as generated Airflow 3 DAGs for scheduled runs. Outputs are canonical MCAP episodes carrying their provenance, backed by a Parquet catalog queried with DuckDB SQL to write a version pinned manifest, and LeRobot Dataset v3 repositories import with hflow import lerobot. The project is pre-v1, with the core lifecycle working end to end.

This targets the unglamorous half of the data problem, where cameras freeze, streams drift out of sync or duplicate recordings enter the corpus. It is the tooling counterpart to the robot data bottleneck, and it operates on the format in rosbag2 recording and replay.

VkVIO accelerates visual-inertial odometry without CUDA

A TU Munich paper (arXiv, posted 24 September, Ole Hoffmann, Mateo de Mayo, Daniel Cremers) presents VkVIO, which the authors describe as the first cross platform GPU accelerated visual-inertial odometry method. Earlier work in this space used CUDA, which ties acceleration to a single vendor. VkVIO uses Vulkan and keeps the causal estimates that real time operation requires.

They deployed it on a workstation, a laptop and an inexpensive single board computer, and report it outperforming CUDA based systems on the same hardware while holding state of the art accuracy. State estimation sets the latency budget for everything above it, so a vendor neutral implementation changes which devices are candidates, alongside the hardware covered in Jetson Orin Nano versus Orin NX.

The robotaxi bottleneck is the depot, not the driver (analysis)

The Robot Report published a guest column by Dan Keene arguing that servicing, not driving, gates autonomous fleet scale. The numbers he cites: Waymo has logged more than 220 million autonomous miles, reports more than 94% fewer serious injury crashes than human drivers, gives about half a million paid rides a week and operates in 14 cities, while Uber serves more than 15,000 cities and did 3.9 billion trips last quarter. A reset, meaning clean, charge and inspect, takes roughly two hours and can involve an empty drive of up to 15 miles to a central depot, and he cites California data showing that of 86 million Waymo miles reported through 2025, only 54% carried a passenger.

Treat the conclusion with the disclosure the column carries: Keene’s own company, Aseon Labs, is building robotic charging, cleaning and inspection stations. The developer relevance survives it, because cleaning and inspection are manipulation problems, and the same servicing gap applies to warehouses and to any humanoid fleet you scale past a pilot.

Quick Bites

• Amazon will invest more than $100 million in a 585,000 square foot assembly plant in Greenwood, Indiana, to assemble automated equipment for its fulfillment centers, including robotic welding and automated powder coating, with 300 manufacturing and engineering jobs expected by the end of 2028. Assembly Magazine.
• ROBOTIS published Mission Canvas, an open source workflow that lets you define waypoints on a map and attach a Task to each location so navigation and robot actions execute as one mission, with tasks able to mix deterministic actions and AI or VLA behaviors. Discourse, GitHub.
• TinyCVIO runs visual-inertial odometry for nanodrones on a dual core microcontroller with 520 kB of SRAM, tracking LED constellations at 29.2 FPS and reporting 3.5 to 3.7 cm mean absolute trajectory error and 15.7 to 16.3 ms estimate latency across nine Crazyflie flights. arXiv.
• The Robot Report podcast episode 263 interviews Sai Vemprala, CTO of General Robotics and previously a senior researcher at Microsoft Research, on the company’s GRID architecture and a modular approach to robot intelligence. The Robot Report.