Intrinsic Core, AWS stack, Atlas hands

Yesterday’s roundup covered the Perseverance record and the Autoware research center. Today: two open source stacks above ROS 2, a benchmark that measures models where demos never go, the reasoning behind Atlas’s new hand, and a buyer for trade data.

Intrinsic opens the stack it sells

Intrinsic released Intrinsic Core at ROSCon 2026 in Toronto, an Apache 2.0 set of ROS compatible capabilities lifted from its commercial industrial platform (The Robot Report, GitHub).

intrinsic_runtime ships as a native k3s environment, so process lifecycle, event scheduling and application state come preconfigured on ROS 2. intrinsic_control, or ICON, is the real time layer: kinematics and trajectory interpolation in a deterministic loop that switches controllers inside one cycle on live sensor feedback, behind a hardware abstraction layer meant to let you swap arms, grippers and fieldbus I/O without rewriting drivers. Around it sit a blended motion planner, 6-DoF pose estimation through NVIDIA FoundationPose, Gazebo services, calibration and ROS drivers.

“What we’re sharing this week is the core of our own software stack,” Brian Gerkey, Intrinsic’s CTO and a co-founder of Open Robotics, told The Robot Report.

That is the gap. ROS 2 gives you topics, services, actions and a build system, and leaves the deterministic loop, the hardware abstraction and the deployment plumbing to every team. Watch which arm and gripper vendors publish ICON HAL plugins next quarter; that list decides whether this becomes infrastructure or a demo.

AWS packages the training loop into one toolchain

AWS launched an open source Physical AI Toolchain that stitches its services to NVIDIA’s robotics software across synthetic data generation, training, simulation, edge deployment and feedback to the cloud (The Robot Report, GitHub). The sample repository is Apache 2.0 and dates from July, so this is packaging rather than new code.

Recordings in Zarr, ROS bag or CSV form convert to LeRobot v2 datasets on S3. Imitation learning fine tunes GR00T N1.6 through SageMaker, reinforcement learning runs in Isaac Lab at 4096 parallel environments on one GPU, Cosmos 3 Predict generates rollouts, Cosmos Transfer 2.5 restyles footage, and a LoRA recipe covers DreamZero, a 14B world action model. OSMO 6.3 on EKS orchestrates cloud and edge, models reach Jetson Thor or AGX through Greengrass, and a Strands Agents layer drives robots in natural language.

“The learnings that get captured locally need to be flown back into the cloud,” Sri Elaprolu, director of frontier AI science and engineering at AWS, told The Robot Report. He cites Amazon’s fleet of more than one million robots, and Telexistence’s convenience store humanoids, of which he says more than 300 are deployed, as the harder case.

The standardisation signal is the LeRobot v2 ingest. Data formats are the quiet lock in of this field, and shipping converters into a community format makes datasets portable. The caution is weight: EKS, GPU quotas and a dozen services is a lot for a small team, so most will lift the ingest and evaluation stages first, where the data bottleneck lives.

A benchmark that puts models inside the control loop

Vals AI published Robo-bench, a held out evaluation of general purpose agents driving simulated robots with only the datasheet, kinematic model, sensors and cameras a designer would hand over (Vals AI). There are 24 tasks: 13 on a 7-DoF Franka Panda arm with no motion planner, and 11 on a 29-DoF Unitree G1 humanoid with two Dex3 hands and no balance controller, so the agent commands every motor including the legs. Both run in MuJoCo under the robots’ real motor limits and a 50 Hz command ceiling.

The scoring design is the point. Each episode is three minutes of wall clock time and physics keeps running while the model thinks, so deliberation costs motion. Balance needs correcting roughly once a second, far faster than a model turn, which means a successful agent has to write code that closes the loop itself.

Top accuracy is 28.33 percent for GPT-6.1 Sol, at an estimated 4.97 dollars per run and 79.89 seconds to first move. GPT-6 Astra reaches 23.33 percent, Claude Opus 5.5 reaches 14.58 percent, Gemini 4 Argon reaches 0.83 percent and Grok 4.7 reaches 0.21 percent. On the public balance scale task, Sol wrote its own controller and carried labelled weights until the beam held level; Opus came closest of the rest and still called the task done before the beam held. Three tasks are public, 21 stay private.

Google reported 87 percent on ERQA for Gemini 4 Argon, a benchmark over images, and the same model sits near the bottom here. Answering questions about a scene and holding a 29-DoF body upright under 50 Hz commands are different problems, which is why the propose and dispose split exists.

Atlas trades a finger for 13 degrees of freedom

Boston Dynamics published the engineering rationale for the hand on its new Atlas: 13 degrees of freedom across four fingers, up from seven on the previous generation (Boston Dynamics). The thumb has four DoF, every other finger three so it can splay and place its fingertip independently, and the hand runs on a single actuator type with no cables crossing joints.

Three decisions stand out. Direct actuation that is backdrivable and transparent, so forces the environment applies show up in proprioception and impacts push the joint aside instead of wrecking a gearbox. Dense pressure sensing across fingertips and palm, letting the controller act on small contact signals. And no pinky: three extra degrees of freedom did not pay for the dexterity gained, a call the team tested partly by taping a finger out of use for a day.

The hand is sized like a large human hand to keep the gap to human demonstration data small, and Boston Dynamics says it is still strong enough to carry a 100 lb or heavier loaded minifridge. The claim that matters for learned policies is simulability: friction, backlash and contact dynamics modelled well enough for sim to real reinforcement learning (simulation stack). It is rated for triggered tool grasps, including drills and welding torches.

HUMXN buys the data robots are missing

HUMXN Research Labs launched a programme to acquire and license 50 million dollars of real world data from skilled trades, farming and manufacturing (Newsfile via Markets Insider, dated 13 October). It wants service records, repair histories, job site photos, manuals and training videos, pays about 20 dollars per hour of approved wearable footage, and says individual deals can pass six figures.

Separately it is subsidising plumbing, electrical, HVAC, duct cleaning and carpet cleaning for homeowners in Minneapolis, Chicago and Miami, in exchange for letting a technician wearing capture gear work inside the house (Tom’s Hardware). Three climates, many house layouts, and tasks no lab can stage convincingly.

“The next generation of AI is going to need more than text and images from the internet,” founder Thomas Unise said. Home repair is a long tail of physical procedures in cluttered, non standard environments, exactly the distribution that robot data pipelines cannot synthesise honestly. Watch whether the datasets ship with the action labels and calibration needed for policy training, since video alone teaches appearance rather than control.

Quick Bites

• ros2_pulse 0.6.0 counts rclpy receives by hooking the rmw_take tracepoint rather than only callback_start, and reports loaned message rates per topic. It exports Prometheus metrics and OTLP on port 9464, and installs from apt on Humble, Jazzy and Kilted (ROS Discourse, GitHub).
• An RFC in gz_ros2_control proposes shared simulation sessions across behavioural assertions that check joint position, velocity and effort responses instead of process and topic existence, with unified semantics from Humble up to Rolling (ROS Discourse).
• A Philadelphia woman was arrested after repeatedly blocking an Uber Eats delivery robot near Reading Terminal Market, telling a local station she was protesting machines replacing delivery workers. Police said she ignored commands to move (6abc).