Atlas hand drops the pinky

Today: Boston Dynamics rebuilds the Atlas hand around manufacturing and simulation, ROS 2 gets a way to declare a node interface in source, and two warehouse robotics companies show what deployment at scale actually looks like.

Boston Dynamics bets four fingers and 13 degrees of freedom

Boston Dynamics unveiled a new hand for its Atlas humanoid on 1 October, and it drops the pinky. The design has four fingers, 13 degrees of freedom, and direct actuation at every joint, up from the seven degrees of freedom in the previous three-fingered hand. The company’s stated goal was not maximum dexterity but a hand that is reliable, repairable and cheap enough to build in the tens or hundreds of thousands.

The trade-offs are explicit. Dylan Thrush, a mechanical engineer on the Atlas team, said the team decided the extra dexterity from a pinky was not worth three more actuators, the added size and the power draw. Before the design was frozen, chief product and technology officer Zachary Jackowski asked the team to tape their pinky to their ring finger for a day and report what they could no longer do. Their conclusion, reported by IEEE Spectrum and The Robot Report, was that the hand did not need it.

Alberto Rodriguez, director of robot behavior for Atlas, frames hands as a ruthless design trade-off, arguing that many humanlike designs give up reliability and manufacturability to look like the real thing. The Atlas thumb carries four degrees of freedom and each other finger three, with fingertips and palm covered by dense pressure tactile sensors.

The developer-relevant part is the simulation story. Boston Dynamics says it built the hand for sim-to-real reinforcement learning: rigid direct drive and a backdrivable transmission, plus control work to compensate for cogging and friction, make the hand cheap to simulate with high dynamic fidelity. That in turn lets RL policies train against randomized motor torque, friction, object geometry and disturbances. The company reports early sim-to-real results on dynamic tasks, though it has published no benchmark numbers. This is the counterpoint to the humanoid hand problems seen elsewhere, where hardware complexity outran the ability to control it (Optimus and the limits of robot hands). Watch whether other humanoid makers follow the good-enough-and-manufacturable path or keep chasing humanlike form.

NoDL lets a ROS 2 node declare its own interface

At ROSCon 2026, Alistair English and Emerson Knapp presented NoDL, the Node Definition Language, and the working group has now opened it for community use. NoDL is a YAML schema and toolkit that describes a ROS 2 node interface: parameters, publishers, subscriptions, service servers and clients, and action servers and clients, plus the QoS profile for each endpoint. The project is Apache 2.0 licensed and lives at github.com/ros-tooling/nodl, with documentation at nodl.readthedocs.io.

The pitch, described in the ROSCon recap, is that a node should be a thin adapter and the interface should be a declared, checkable contract. You can scrape a document from a running node with ros2 nodl describe <node_name>, then use it to check that the node conforms to the spec, generate documentation, generate C++ or Python base classes, and map an application graph statically without launching it. If you already think of ROS 2 interfaces in terms of topics versus services versus actions and tune QoS policies by hand, NoDL is the declarative version of that work. The working group also plans to absorb generate_parameter_library into the project, and meets fortnightly.

Destro AI sells the orchestration layer, not the robot

Destro AI came out of stealth with an $8 million seed round led by Base10 Partners and Bonfire Ventures, and its pitch is unusual: the best way to deploy a robot is often not to build one. TechCrunch reports that Destro runs a cross-docking workflow at Yusen Logistics where humans, carts and trucks are all directed by its Mothership operating system, while three Miva Robotics cart movers run on Destro’s Vision OS, built on open-weight vision-language-action models.

The scale-up is concrete: the initial pilot expands to 26 robots, with a second pilot of 17 robots at a Southern California facility. The interesting claim is architectural, not mechanical. Destro argues the hard part is orchestration across mobile robots, fixed automation, warehouse systems and people, rather than the manipulation policy itself, and that it can copy the workflow across warehouses. Founder Manthan Pawar concedes the model may not hold for dexterous manipulation, where general robots and open models are not yet reliable. For developers, the takeaway is that the integration layer is where a lot of near-term value sits, on top of the open action models now shipping.

Maven Robotics reaches 3 meters with 30 kilograms

Maven Robotics has raised $100 million since 2024 and runs its mobile manipulators 16 hours a day, five days a week, with up to 99 percent uptime, according to CEO Hamza Derbas. The robot is a wheeled base with an arm that reaches the back of a 48 inch pallet and up to 3 meters high while carrying 30 kilogram payloads, a combination of large dynamic reach and a footprint small enough to pass through doors.

Maven started from the work, not the morphology: mixed-case palletizing and tote handling first, with assembly and fabrication later. On tooling, Derbas says off-the-shelf grippers underperformed, so the company iterated its own and settled on a miniaturized vacuum gripper plus a two-finger parallel gripper on a modular end-of-arm interface. It also runs a shared learning fabric it calls the Maven Network. The line to watch is whether reliability, rather than day-one generality, stays the selling point, the same argument behind the robot data bottleneck.

Agility and FORT formalize the offboard safety layer

Agility Robotics and FORT Robotics signed a memorandum of understanding on 1 October to extend Digit 5 safety architecture beyond the robot itself, per the Agility press release. FORT has supplied Digit hardware for years, including the pendant sled that holds the robot tablet with emergency stop and enabling-device functions. The new work is an off-robot interface, the Offboard Safety Bridge, connecting Digit to external safety systems.

Agility positions this as a third layer alongside Safe Human Detection and the independent safety controller in Digit 5 cooperative safety. The practical value for integrators is a redundant manual override and continuity of safety coverage as deployments grow, and both companies say the partnership will cover joint hardware development, regulatory compliance and deployment support.

An open source desktop robot on Arduino UNO Q

If you want a cheap way into physical AI, the JBR-001 is an open source, 3D printable desktop companion robot built around the Arduino UNO Q. It uses Modulino Distance, Motor and Buzzer modules plus an LED matrix, and ships with Arduino App Lab examples for head and arm servo motion, distance sensing and sound. The UNO Q pairs a microcontroller with a Linux side for vision and edge AI, the same split robot builders keep making: real time control in one place, perception in another. Build instructions and code are on the project page.

Quick Bites

• ROS 2 Jazzy has a sync planned for Monday 5 October, with 472 packages ready and no regressions reported (Open Robotics Discourse).
• The ROS PMC accepted REP 158, defining OpenUSD conventions for simulation asset interoperability, and named new committer candidates after ROSCon (ROS PMC minutes).
• HumanoidTTT, an arXiv paper on reusing validated humanoid motions at test time, reports zero unsafe reuse and a 16.4 times end-to-end speedup over fresh generation, with code released (arXiv:2610.00198).
• Fig released RIDGE, item-level evaluation across five web and physical benchmarks, finding no clear performance leader and that model updates reshape which tasks a model solves even when the average barely moves (Fig).