Isaac ROS 5.0 lands at ROSCon

Today: ROSCon week in Toronto brings two large open source releases, a serving runtime that shortens vision-language-action inference on edge hardware, a bimanual diffusion policy that predicts geometry while it acts, one engine line where vision replaced fixture precision, and a wheeled single arm built for homes rather than demo reels.

NVIDIA puts agents inside Isaac ROS 5.0

NVIDIA released Isaac ROS 5.0 at ROSCon 2026 in Toronto, the GPU-accelerated package collection that sits on top of ROS. The release adds support for ROS Lyrical and Ubuntu 24.04, and NVIDIA says it worked with the Open Source Robotics Alliance to contribute a standard data-handling interface to ROS Lyrical, with CUDA as the working example for GPU acceleration. That interface is arguably the durable part: it gives one consistent way to move robot data across CPU and accelerator, and it landed in the ROS distribution rather than in a single vendor’s stack.

The agentic framing is both the pitch and the substance. New Isaac skills for setup and manipulation give developers and coding agents reusable workflows, and the documentation is written to be agent-readable. A FoundationStereo fine-tuning skill lets an agent adapt a stereo depth model to a specific camera, environment and application. FoundationPose, the 6D object pose model, now has an agent-ready inference library that NVIDIA says makes pose estimation and tracking up to 5.5 times faster, and pick and place is available as a standalone skill that runs outside Isaac ROS. Those speed and support figures are NVIDIA’s published claims.

The ecosystem list says as much as the feature list. RealSense sponsors AgenticROS, an open source project that connects Isaac ROS to NVIDIA Nemotron models and NemoClaw blueprints, and is optimizing its newest stereo cameras for Isaac ROS and Jetson Thor. Ekumen reports that isaac_ros_cumotion on a GPU maps a collision-free path for a warehouse arm in roughly 2 to 5 milliseconds. Seeed Studio pairs Isaac ROS with its reBot Arm on Jetson Thor, Magna pairs it with Isaac Sim for hardware in the loop testing, Universal Robots has built it into its AI Accelerator SDK, ROBOTIS into its AI Worker, and Mentee Robotics uses it as the perception backbone of MenteeBot. Support runs from Jetson Orin Nano to Jetson Thor, which is where choosing between Orin Nano and Orin NX becomes a platform decision rather than a hobby question. Background: what NVIDIA Isaac Sim is.

Source: NVIDIA, The Robot Report.

Intrinsic open sources the core of its robot platform

Intrinsic published Intrinsic Core under the Apache 2.0 license at ROSCon 2026 in Toronto: a local runtime, SDK and hardware agnostic real-time control framework for industrial robotics. The company, which joined Google this year, built the release from capabilities it had already run in real manufacturing deployments instead of writing a fresh demo stack.

The module list reads like the layer ROS 2 leaves to integrators. intrinsic_runtime is a k3s containerized execution engine. intrinsic_control, or ICON, is a deterministic control loop that can switch controllers within a single cycle based on live sensor feedback, sitting on a hardware abstraction layer that swaps arms, grippers and fieldbus I/O without driver rewrites. Motion planning covers both Cartesian tasks and wide-envelope joint-space moves; perception embeds NVIDIA FoundationPose for 6D part pose; inference serves models locally. Requirements are Ubuntu 24.04 LTS and ROS 2 Lyrical Luth. The repository also ships an open machine tending reference design, MoveIt 2 integration, a shared memory transport, ROS camera drivers and malloc-guard, a library for catching dynamic heap allocations in real-time threads.

Brian Gerkey, Intrinsic’s CTO and a co-founder and board chair of Open Robotics, told The Robot Report that this is the core of Intrinsic’s own software stack and not code dropped over the wall, and Geoffrey Biggs, CTO of the Open Source Robotics Foundation, described it as the direction robotics development is heading. For teams building cells today, the practical question is whether a permissively licensed real-time control loop plus a hardware abstraction layer removes enough integration work to be worth adopting over a hand-rolled stack.

Source: GitHub, The Robot Report.

A serving runtime that shortens VLA inference on Jetson Thor

InstinctFlash is an open source serving runtime for robotics models that appeared on Hacker News this week. The repository has been public since a full source release on 15 September, covering eight model families that include pi0.5, GR00T N1.7, LingBot-VA and LingBot-VLA, NVIDIA’s Cosmos3 Edge and Nano policies and DreamZero. RTX 4090 and RTX 5090 support followed on 16 and 17 September, using the same Runtime API as Jetson Thor.

The interesting numbers are the project’s own measurements on Jetson Thor, and they deserve reading line by line. On matched sampling schedules, pi0.5 with FP8 drops from 408.58 ms to 51.85 ms median prediction latency, a 7.88x reduction, LingBot-VA at 25 views and 50 actions goes from 15506.32 ms to 2891.74 ms (5.36x), Cosmos3 Edge DROID from 3393.78 ms to 1048.01 ms (3.24x), and GR00T N1.7 in bit exact mode from 139.50 ms to 117.30 ms (1.19x). The headline 33.78x figure combines FP8, fewer sampling steps and a schedule change from 25 views and 50 actions down to 2 views and 4 actions, so the same-schedule rows are the honest comparison. Precision defaults to native with a bit exact ceiling, and FP8 is opt in.

The serving path speaks msgpack over websocket in the protocol the pi0 and openpi ecosystem already uses, so existing robot side clients connect without changes. If latencies like these hold on your hardware, policies that were too slow for a tight control loop become plausible, which is the central tradeoff behind why VLA models matter.

Source: GitHub.

JAMB predicts geometry while a bimanual policy acts

arXiv paper 2609.25322, submitted 21 September by Chuyang Xiao, Peilin Meng and David Held, proposes JAMB, a diffusion policy that jointly denoises bimanual actions and future 3D point tracks inside one Transformer. Each hypothesis informs the other during denoising, and the multimodal representations are grounded in a shared spatiotemporal coordinate system. The motivation is specific to two arms: motion of either arm changes the shared scene, and most diffusion policies generate actions without modeling that geometric consequence, while predictive variants use future state only as auxiliary supervision.

The results are reported across 16 simulation tasks in RoboTwin 2.0, with an 83.4 percent average success rate, 23.9 percentage points above the strongest baseline. On three real-world tasks the authors report gains of 50.0 and 21.2 percentage points over action-only and auxiliary geometry prediction methods, plus better generalization to cluttered scenes and out-of-distribution backgrounds. If you already run a diffusion policy on a dual-arm cell, joint track prediction is the kind of change that can be tested without replacing the controller.

Source: arXiv.

Vision guidance replaced fixture precision at Stellantis

Assembly Magazine reports that at Stellantis’ Trenton, Michigan engine plant, a six-axis robot tightening bolts was missing 2.55 percent of them because the lift-and-locate station could not position each 3.2 or 3.6 liter V6 engine repeatably enough, which put a person at the end of the line to re-tighten whatever was missed. Stellantis engineers deployed Inbolt’s guidance system: a 3D camera mounted on the robot captures the station in real time, a model pretrained and then fine-tuned on CAD data locates each fastening point, and the robot path is corrected on the fly while parts are still moving. The article notes the system was installed without infrastructure changes and needs no fixed lighting or controlled environment.

This is the industrial version of the perception argument the Isaac ROS and Intrinsic releases make. A fixed fixture buys repeatability with infrastructure, cycle time and wear; a vision model that generalizes across part geometry buys flexibility instead. The 2.55 percent miss rate is the plant’s figure as reported by Assembly, and the measurable outcome to ask vendors for is exactly that kind of number on a live line.

Source: Assembly Magazine.

Stretch 4 makes the case for one arm on wheels

Hello Robot’s CEO and co-founder Aaron Edsinger brought Stretch 4 to the Real World AI Stage at TechCrunch Disrupt 2026, TechCrunch reports, with a session framed around usefulness rather than a demo reel. The robot is a wheeled base with a single telescoping arm, and it launched in May. Hello Robot’s published specifications list 55 cm of reach plus a 6 cm wrist, 2.5 kg payload with the arm extended and 4 kg retracted, 46 kg total with 33 kg of ballast removable for transport, 8 hours of runtime at light CPU load, a 160 cm height and 45 cm footprint diameter, an omnidirectional base with self charging, contact sensitive design for working near people, and an open source ROS 2 and Python SDK with the sensors calibrated out of the box. Reference demos cover mapping, navigation, 3D SLAM, data collection and VLM grasping.

While much of the industry optimizes humanoids, Hello Robot trades reach, dexterity and two arms for a 45 cm footprint, a stable base and a simpler control problem, and the deployment targets are people with severe mobility impairments who use the robot to eat, handle everyday objects or close blinds. Those specifications are the manufacturer’s, not independent measurements. The developer relevant question is whether the ROS 2 SDK and the shipped navigation and manipulation demos shorten the distance from a lab policy to a working assistive application.

Source: Hello Robot, TechCrunch.

Quick Bites

• The Isaac ROS team posted a walkthrough of collision-aware motion planning with Isaac ROS cuMotion on the ROS Discourse projects board, the same GPU planner Ekumen reports running in single-digit milliseconds for a warehouse arm. ROS Discourse
• The Maritime Dual-Use Hackathon runs in San Francisco from 16 to 18 October with a 48 hour build, a Software Defined Vessel track plus piping layout optimization and OT cybersecurity challenges, organized by Silicon Mtn, Fincantieri, Fincantieri Marine Group and Mind the Bridge, with Qualcomm and Arduino as technology sponsors and teams of up to five. ROS Discourse
• arXiv:2609.25274 learns the interaction manager of a collaborative robot instead of hand coding it, training a multimodal reinforcement learning policy on a simulator seeded with human data for a home robot helping a user locate objects, and validates it in a human study. arXiv
• A useful operations read: Curiosity project scientist Ashwin Vasavada describes the JPL rover shift, where a few dozen people get three hours to assess rover health and new science data, then four hours to assemble tomorrow’s commands inside the rover’s energy and time budget before sunrise on Mars. Robohub