PickNik to Qualcomm, Intrinsic goes open

Today: the ownership of the most used manipulation framework in ROS changes hands, Google’s Intrinsic puts its industrial stack on GitHub, Boston Dynamics opens a place to teach Atlas on real car parts, and a Stanford group asks whether you still need a learned policy at all.

Qualcomm agrees to acquire PickNik and keep MoveIt open source

Qualcomm Technologies announced an agreement to acquire PickNik, the Boulder company that stewards MoveIt, and said it plans to keep MoveIt 1 and 2 open source under their existing license with community driven roadmaps. PickNik formed in 2015 and its team had started on MoveIt inside the ROS project at Willow Garage. It rebranded MoveIt Studio to MoveIt Pro in 2024 and shipped MoveIt Pro 9.0 this year with perception to motion capabilities and a training data collection system.

Qualcomm frames the deal as a path from models to planning, manipulation and real time control on its Dragonwing robotics and Arduino VENTUNO Q platforms, after acquiring Arduino in 2025. Qualcomm says it will keep supporting MoveIt on third party hardware, and no amount was disclosed.

For anyone maintaining a MoveIt configuration or a MoveIt Pro deployment, the roadmap owner just changed. MoveIt is where much of the arm planning in ROS 2 actually happens, on top of the middleware covered in what is ROS 2. Watch whether the permissive license holds and whether Dragonwing becomes the best tested compute target.

Intrinsic open sources its industrial robotics stack

Google’s Intrinsic announced Intrinsic Core at ROSCon 2026 in Toronto on 22 September: ROS compatible capabilities under Apache 2.0, on GitHub, described as the same parts the company uses in its own manufacturing deployments.

The modules matter individually. intrinsic_control is a hardware agnostic real time control engine that switches controllers inside one cycle from live sensor feedback, behind a hardware abstraction layer so arms and grippers can be swapped without rewriting drivers. intrinsic_motion_planning blends collision free path segments with differing velocity profiles. intrinsic_perception ships with built in support for NVIDIA FoundationPose, for six degrees of freedom pose estimation of unfixtured parts. intrinsic_runtime is a local execution engine packaged as a k3s containerized environment on top of ROS 2. Simulation services run on Gazebo, and Intrinsic also published a machine tending reference design.

The claim to test is an integrated local stack a factory can run on premises, with a HAL instead of per vendor drivers. That removes a large part of the integration tax that keeps research code off production lines. See what is NVIDIA Isaac Sim for the other simulation route.

Boston Dynamics opens a training center for Atlas

Boston Dynamics opened the Robotics Metaplant Application Center near Savannah, Georgia, a testbed for integrating Atlas into Hyundai Motor Group’s automotive manufacturing. The group plans to expand Atlas deployment beginning with 25,000 units at Hyundai Motor and Kia plants worldwide over the next few years.

Atlas is learning from teleoperation, reinforcement learning in simulation, and Universal Manipulation Interface devices. The revealing part is what the site changes besides the robot: the company says it is learning how to modify fixtures and packaging to make them easier for humanoids, how to collect data inside a production environment, and how to build support processes around the machines. If a task only succeeds after the fixture is reworked, the cost lands on whoever integrates the cell.

A frontier VLM drives a humanoid with no learned VLA in between

A Stanford group (Gio Huh, Cayden Gu, Takara E. Truong, C. Karen Liu, Guy Tevet) presented HomeBody, which questions the usual three part autonomy stack: a System 2 vision language model for reasoning, a learned System 1 vision language action model for commands, and a System 0 controller for whole body motion. HomeBody drops the learned VLA and has the VLM orchestrate a library of composable motor skills directly, backed by persistent spatial memory.

In a previously unseen kitchen, a Unitree G1 guided by GPT Astra tidies the room, then retrieves a remembered object from a request such as “I forgot my medicine,” with no environment specific training data and no additional policy learning. Astra also acts as the Real2Sim agent, building an Isaac Sim digital twin from the robot’s own exploration data, which combines phone video, RealSense D435i observations, LiDAR with SLAM and joint poses, so the model can reason about places it cannot currently see.

Read it as an argument about where value accrues. If a frontier VLM can sequence skills, the work moves into the skill library, the spatial memory and the feedback loop that lets the model revise after a failure, not into another policy. These are the authors’ own results and the project page notes robot poses are illustrative. Compare the scaling route in why VLA models matter.

An open serving runtime attacks robot inference latency

InstinctFlash released its full source on 15 September, added RTX 4090 support on 16 September and RTX 5090 on 17 September, with the same Runtime API used on Jetson Thor. It covers eight robotics model families, including pi05, GR00T N1.7 and several LingBot vision language action variants, with Python and WebSocket serving.

On Jetson Thor, with prediction p50 latency measured on 15 September, the project reports pi05 falling from 408.58 ms in PyTorch to 51.85 ms with FP8 (7.88x), LingBot-VLA-4B from 624.22 to 221.53 ms (2.82x) and GR00T N1.7 from 139.50 to 117.30 ms (1.19x). The headline 33.78x figure compares a 25V/50A schedule against a 2V/4A one, so it is not the same measurement, and the authors say they saw no loss in task performance in their own real robot tests.

Treat all of it as vendor benchmarking that nobody has replicated. It still matters because action chunk latency sets how fast a policy closes its loop, and moving a 400 ms prediction to 50 ms changes which tasks are viable on the same silicon. Hardware choice runs alongside, as in Jetson Orin Nano versus Orin NX.

ABB moves robot programming onto your own tablet

ABB Robotics launched E-Device, a safety interface that connects a Windows tablet or PC to the FlexPendant connector on an OmniCore controller over USB C and runs the OmniCore user interface app. Engineers simulate offline in RobotStudio, then carry the same device to the line for commissioning, with AppStudio available for no-code screens. The safety rated parts stay in the interface: a physical emergency stop button and a three-position enabling switch. ABB says it complies with ISO 10218:2025 and is CE and UL certified.

Moving the operator interface onto commodity hardware makes certification a dongle question rather than a panel question.

Universal Robots redesigns its platform around the tool flange

Universal Robots introduced Gen 7 at IMTS in Chicago on 14 September, anchored by the PolyScope X operating system with a reengineered control box, teach pendant and tool flange. The new flange delivers data, power and safety directly to the tool, which the company says lets vision and sensing for AI models run in real time with the robot. The CB7 controller ships in standalone and cabinet versions and stays compatible with e-Series and UR Series force and power limited arms.

Where the sensing sits is the decision with consequences. If power and data reach the flange, cameras, grippers and force sensing stop being separate integration projects and become configuration.

Quick Bites

• ros2_unbag reached 1.3.0 and is now in the official ROS 2 buildfarm, so installation is sudo apt install ros-<distro>-unbag then ros2 unbag, with no clone or source build. It exports images and video, point clouds, CSV, JSON and YAML, with topic filtering, parallel extraction and custom processors. Discourse, GitHub. Pairs with rosbag2 recording and replay.
• An experimental CPU shared memory backend for rosidl::Buffer shares the backing storage of variable length payloads such as images across processes, so bytes skip the normal message path and Python subscribers read a NumPy view without another copy. The author notes Fast DDS data sharing requires bounded types. Discourse, GitHub.
• Figure and Nscale agreed to deploy up to 100,000 NVIDIA GPUs on the Vera Rubin platform, targeting initial deployment from the second half of 2027 in Barstow, Texas. Figure says its Helix models are bound by data and compute. Figure.
• Feather Robotics described a modular humanoid platform for developers, with configurable hardware and software able to run models from NVIDIA, Skild AI or Physical Intelligence. The company says its robots work as cooks and clean science labs. TechCrunch.