Intrinsic open-sources its robot runtime

Intrinsic released the runtime it uses for manufacturing cells as Apache 2.0 code, Agility Robotics moved within days of a Nasdaq listing, and Tutor Intelligence raised again on the argument that deployment is the training loop. Also: a quadcopter that chased a runner through a forest with no GPS, and Gazebo learns to namespace a model.

Intrinsic open-sources an industrial runtime for ROS 2

Intrinsic, the Alphabet-owned industrial robotics software company, published Intrinsic Core on GitHub under Apache 2.0 after announcing it at ROSCon (ROS Discourse, repository). The repository was created on 8 September 2026, carries 638 stars, and ships frequent tagged releases, the newest being 20261022.0.

It is a preconfigured local runtime in a k3s container plus the pieces a workcell needs. intrinsic_control, or ICON, combines kinematics solvers, trajectory interpolation and a deterministic control loop that can switch controllers within a single cycle based on live sensor feedback, behind a hardware abstraction layer. intrinsic_motion_planning covers Cartesian and C-space planning, blending segments with different velocity and acceleration limits into one trajectory. intrinsic_perception standardizes camera and point cloud interfaces and ships NVIDIA FoundationPose for 6-DoF pose estimation of un-fixtured parts, served locally by intrinsic_inference.

The build targets C++20 with Bazel on Ubuntu 24.04 LTS and ROS 2 Lyrical Luth, with 22.04 listed as supported. Existing nodes keep working through standard workspaces and message definitions, and a separate intrinsic-moveit repository bridges to MoveIt 2. The reference cell is the Open Machine Tending Solution for CNC tending. Intrinsic says community members have already adapted it to insert car wheels with a custom gripper on a larger arm, running inference locally on ARM.

Industrial cells usually couple controller, arm and gripper into one vendor decision. A hardware abstraction layer behind a ROS 2 message boundary is the seam that lets a shop swap a gripper without rewriting drivers. Watch whether third-party hardware plugins arrive, and whether the timing claims hold on ordinary industrial PCs. Background: what ROS 2 actually is.

Agility clears the SPAC vote and lists this week

Shareholders of Churchill Capital Corp XI approved the business combination with Agility Robotics on 2 November, with more than 84 percent of votes cast in favour (The Robot Report). The deal should close on 4 November and the combined company is expected to begin trading on Nasdaq as AGLT on 5 November, against a $2.5 billion pre-money equity value.

The capital figure fell. Agility now expects about $465 million in gross proceeds rather than the more than $620 million anticipated in June: roughly $164 million from the trust account, $201 million from a PIPE led by Foxconn, and $100 million in equity raised after the announcement.

Digit runs in 9 customer environments. GXO Logistics, the anchor account since 2023, is expanding from Flowery Branch, Georgia to Statesville, North Carolina and Lebanon, Tennessee, starting in early 2027 on Digit 4. AGCO signed for a Hesston, Kansas plant in 2027, beginning with sorting, stacking and tote handling. Agility has claimed more than $300 million in Digit 5 orders, which is a multi-year order book subject to milestones, not revenue. RoboFab in Salem, Oregon is built for up to 10,000 robots a year, and Digit 5 early access is targeted at the first half of 2027.

A public humanoid company means quarterly disclosure of deployments and revenue against those order claims. That follows the 24.8 percent rise in North American robot orders reported last week. Watch the first 10-Q for the gap between contracted and delivered.

Tutor Intelligence raises $75M on deployment as training loop

Tutor Intelligence, an MIT CSAIL spinout, raised a $75 million Series B led by Andreessen Horowitz, taking total capital to $117 million (The Robot Report). Its Sonny robots are two-armed semi-humanoid machines doing each picking in customer facilities, trained on data from that work with help from remote human “tutors”.

The pricing model is the part worth copying. Tutor charges by the hour or by the pick, with no upfront purchase and no long term contract, and handles deployment and maintenance itself. It says twice as many robots are working in almost four times as many facilities as a year ago, that nearly a third of recent recipients were existing customers adding more, and that it expects more than 1,000 robots under contract by the end of 2027.

Per-pick pricing turns model reliability into a revenue question. A policy that succeeds 90 percent of the time leaves a tenth of the picks to a human, and that residual is what the data flywheel has to shrink. Same logic as the robot data refinery, with a business model attached.

A drone chases a runner through trees without GPS

Aniket Bera, a Purdue computer science professor, posted a 44 second video of a quadcopter following a person through a forest with no GPS, no map and no cloud link, with everything processed on board (TechSpot). The post describes an occlusion prediction step that estimates where the target went while it is hidden behind a trunk.

Provenance is the caveat. The demo exists as a LinkedIn video, the post does not say what the drone runs, and TechSpot noticed the overlays label the motion planner SUPER and the altitude estimate FAST-LIO, both open source projects from the University of Hong Kong’s MaRS Lab. SUPER is the lab’s LiDAR guided drone, published in Science Robotics with flights above 20 meters per second through unknown terrain, and its design files are public (SUPER, FAST-LIO). No paper on the forest flight appears on the Purdue lab’s publications page.

The reproducible part is the stack: LiDAR inertial odometry for state estimation, an open source planner, and a target tracker on top. The claim that every layer was built from scratch sits awkwardly beside those on-screen labels. Watch for a paper or a code release; until then the architecture is the takeaway and the authorship claim is unverified.

Gazebo adds model namespaces and automatic bridges

A Google Summer of Code 2026 project added model namespace support across SDFormat, Gazebo Sim and the ROS integration, removing the main busywork in multi-robot simulation (ROS Discourse). A <namespace> element can be set in the SDF file or when spawning a model at runtime, and supported plugins use it to resolve topic and service names. Nested models resolve by walking up the parent chain and combining relative namespaces until an absolute one is found, and absolute names stay untouched, so existing worlds behave the same.

The second feature is an automated bridge. Instead of one hand written bridge per robot, it discovers interfaces on the ROS 2 and Gazebo sides, matches them by name and compatible message type, and creates the bridges itself. It is optional, enabled with an automated_bridge parameter.

Namespace support is in the DiffDrive plugin and extended to other drive plugins and sensors, with several pull requests still under review. Until it reaches a Gazebo release, multi-robot launch files still need per-robot SDF copies, and plugin authors must opt into namespace resolution or their topics collide. The same scaling problem shows up on the Isaac Sim side.

Quick Bites

• Luxonis opened the OAK 4 FFC, a 70 by 53 mm board with five 4-lane camera ports, 52 TOPS (48 INT8 plus a 4 TOPS FP16 GPU), a 6-core ARM CPU, 8GB of RAM and 128GB of storage (Luxonis, IEEE Spectrum).
• Ford is putting $2 billion into its Louisville Assembly Plant to replace the moving line with a three-branch process that builds vehicle sections simultaneously, targeting 40 percent faster assembly (ASSEMBLY Magazine).
• A University of Bristol team drove electric motors through a shared mechanical backbone and watched muscle-like load-speed behaviour and travelling waves of coordination emerge without central control, published in Royal Society Interface (Robohub).