Boston Dynamics' new CEO and Atlas hand

Today: Boston Dynamics puts an Amazon AI veteran in charge and strips a finger off Atlas, Intrinsic open sources its core robot stack under Apache 2.0, Amazon confirms a warehouse fleet past one million robots, NVIDIA opens the Vera Rubin rack design, ABB adds a compact arm family, and ROS 2 actions reach the browser.

Boston Dynamics names Rohit Prasad CEO

Boston Dynamics named Rohit Prasad chief executive on 6 October, with the appointment taking effect the next day (The Robot Report, Boston Dynamics). Prasad spent 12 years at Amazon, most recently as senior vice president and head scientist for Alexa, and later led the Amazon Nova foundation models for enterprises, after nearly 14 years in machine learning research at Raytheon BBN Technologies.

He is the company’s third CEO, after founder Marc Raibert and Robert Playter, who stepped down in February 2026, with CFO Amanda McMaster holding the role in the interim. His stated plan, productizing intelligent robotic systems for industrial and commercial environments, points at the model and software layer rather than new hardware alone. Spot 5.2 is pitched as embodied AI for the factory floor, and Orbit AIVI inspection now runs on Google Gemini Robotics (Boston Dynamics news). Watch how much of that stack stays closed.

Atlas gets a 13 degree of freedom hand, no pinky

A week earlier Boston Dynamics showed the newest Atlas hand: four fingers, 13 degrees of freedom, directly actuated, each joint driven by one of 13 identical encapsulated actuators with no cables crossing joints (Boston Dynamics, The Robot Report).

The previous hand had seven degrees of freedom. The thumb now carries four, and the other three fingers carry three each. The company says the hand still handles more than 100 lb of payload, with dense pressure tactile sensors across fingertips and palm. Strength was already sufficient, so the team traded some of it for actuation transparency and backdrivability, which feed proprioceptive control. Four fingers remove three actuators, and with them size, power draw and cost, a decision engineers tested by taping a pinky and ring finger together for a day. Keeping the geometry close to a large human hand keeps the cross embodiment gap small for human demonstration data, and lets the hand be modelled in simulation for sim-to-real reinforcement learning. Watch for published manipulation benchmarks rather than demo clips.

Intrinsic open sources its core robot stack

Intrinsic, the robotics software group now inside Google, released Intrinsic Core at ROSCon 2026 in Toronto under Apache 2.0, drawn from what it runs in production manufacturing deployments (Intrinsic, The Robot Report).

The pieces matter more than the announcement. Intrinsic Control is a hardware agnostic real-time control framework that adapts a trajectory mid-motion and swaps arms, grippers and sensors without rewriting drivers. Pose estimation builds on NVIDIA FoundationPose for six degree of freedom part localization, motion and grasp planning generate collision free paths, simulation services run on Gazebo, and the release adds camera calibration plus preconfigured ROS drivers. CTO Brian Gerkey, an Open Robotics co-founder, framed it as the core of the stack Intrinsic itself depends on. Teams that fight executors and callback groups for timing should look at how Intrinsic Control schedules work.

Amazon passes one million warehouse robots

No organization operates more robots than Amazon. The company says more than one million robots now work in its fulfillment network, and it is building extra production lines for them in Austin, Texas and Greenwood, Indiana. The fleet began with the 2012 acquisition of Kiva Systems, which became Amazon Robotics, and Amazon is still the second-largest corporate employer in the world, with more than 1.5 million full time and part time workers (Ars Technica).

Most of those robots are wheeled mobile units. Hercules and Titan carry the tall pod shelves, Proteus uses sensors to work safely around people rather than staying behind fences, and DeepFleet, a set of foundation models, predicts robot congestion in the aisles. Amazon claims its Sequoia system identifies and stores items up to 75 percent faster by coordinating swarms of them. The arms are where the machine learning shows up. Amazon Robotics has single-digit thousands deployed and expects to more than double that fleet this year, per the CFO’s July earnings call. Robin, an eight suction cup arm that sorts mailers out of cluttered piles, replaced hand crafted picking strategies with a learned model trained on more than 394,000 picking actions, which measured 93.73 percent pick success against 92.28 percent for the best traditional algorithm, about 16,500 fewer failed picks across more than a million picks. A later online experiment over two million picks cut the failed pick rate by 20 percent, presented at the International Symposium on Experimental Robotics in July 2025. Sparrow was the first place Amazon put foundation models against catalog diversity, and the team has since built a perception foundation model for products it has never seen plus a generalized action model spanning arms that use suction and finger grasping. Vulcan adds force feedback, a first for the fleet.

Reports from the Strategic Organizing Center in 2021 and 2022 found higher injury rates in Amazon’s robotic warehouses, and George Mason University’s Brian Greenwood notes that collaborative robots stand close to workers and can set the pace of the work. Amazon says robotic maintenance technician roles in its network grew by more than 30 percent. The developer lesson runs through VLA research: models that generalize across embodiments are what make a fleet this size affordable to run.

NVIDIA opens the Vera Rubin rack design

At the OCP Global Summit NVIDIA contributed its Vera Rubin rack and compute tray designs to the Open Compute Project through MGX, extending the third generation reference design with a 1.2 specification (NVIDIA). The spec adds rack scale blind mate liquid cooling, a liquid cooled 5,000A busbar, native 19 inch equipment support and denser racks, and NVIDIA says racks can reach 250 kW. A separate Vera CPU rack targets agent orchestration and tool calling rather than training.

The robotics bridge is indirect but real: robot foundation models, world models and video based policies train on clusters like these, and open rack specifications feed the simulation and data pipeline work teams depend on.

ABB adds a compact IRB 1700 robot family

ABB Robotics expanded its industrial line with the IRB 1700 and IRB 1700ID, compact robots in five variants with payloads from 6 to 12 kilograms on OmniCore controllers (Design World). ABB quotes cycle times up to 20 percent shorter and power consumption up to 15 percent lower than the IRB 1600 they succeed, with service intervals up to 20,000 hours between oil changes, path repeatability of 0.03 millimeters, joint speeds up to 700 degrees per second and a footprint 50 percent smaller. The ID variant routes cables and hoses inside the arm. Targets include material handling, arc welding, deburring, sanding and machine tending, plus AI server rack assembly, and the family works with RobotStudio, AI Robot Trainer and OmniCore EyeMotion. ABB showed it at the China International Industry Fair in Shanghai from 12 to 16 October.

ROS 2 actions reach the browser with rclnodejs 2.3.0

rclnodejs, the ROS 2 client library for Node.js, shipped version 2.3.0 with action support over WebSocket and HTTP/SSE (GitHub release, ROS Discourse). The browser SDK derives TypeScript goal, feedback and result types from generated ROS declarations, and the OpenAPI export now carries action schemas, so a web front end and an agent tool definition can come from one description. An HTTP runtime can expose a running action server on port 9001, sending goals and streaming feedback works over both transports, and cancelling a goal is WebSocket only. Topics, services and SSE streaming landed in earlier releases, all behind the same capability allow-list. For teams that keep a browser dashboard beside a robot this replaces a custom bridge, and it complements the topics, services and actions split the rest of the stack assumes.

Quick Bites

• ROS 2 Humble received its 2026-10-12 sync: 40 new packages and 446 updates, including EasyNav 0.5.0 arriving as binaries after its source release (ROS Discourse).
• Rokoko shipped a new Device SDK under its ROS 2 bridge for Smartgloves II: a sturdier USB stream, firmware with lower power draw, over the air updates planned for the next release, and a pinch detection example covering thumb to index, middle, ring and pinky pinches on both gloves (ROS Discourse).
• DendROS is an MIT licensed wrapper that colors ROS 2 terminal output per node group from a dendROS.yaml file and adds a scrollable, searchable viewer for ros2 launch, with no dependency beyond PyYAML (GitHub).
• Teradyne made a strategic investment in Bright Machines and will integrate its robotics and test systems into Bright Machines’ AI data center manufacturing platform, covering robotic assembly, test equipment loading and autonomous material movement (The Robot Report).
• Epson Robots introduced the AX6, a six axis collaborative arm with a 900 mm reach and a 6 kg payload, paired with an RC-A1010 controller that supports a no-code interface and a Python development environment; Epson says it has sold more than 250,000 robots as of April 2026 (The Robot Report).