Zero-copy ROS 2 and a $10 tactile skin

Today: ROS 2 gets a zero-copy path for GPU payloads and an experimental way to survive a supervisor restart, a free tool turns a robot description into a buildable workspace, and low-cost tactile skins make human demonstrations reusable across embodiments.

NVIDIA takes the CPU out of the ROS 2 GPU path

NVIDIA contributed a CUDA buffer backend to ROS 2 Lyrical, built on the upstream rosidl::Buffer<uint8_t> abstraction. In Lyrical, variable length primitive array fields such as uint8[] are generated as a rosidl::Buffer<uint8_t> that behaves like the std::vector<uint8_t> interface existing code expects. The CUDA backend implements that storage with CUDA Virtual Memory Management, so payloads move between co-located nodes without serialization or host copies when both endpoints meet the requirements: same host, same CUDA device, same Linux user and a supported RMW implementation such as rmw_fastrtps_cpp or rmw_zenoh_cpp. Anything else falls back to the CPU path, which is what makes it deployable rather than an all or nothing switch (NVIDIA Technical Blog).

NVIDIA says every node in Isaac ROS 5.0 now uses the backend. The post migrates a real node, the Depth Anything 3 TensorRT node, which takes an image, runs metric depth inference and publishes a float depth image. Its kernel was never the problem: the CPU-backed ROS boundary around a GPU native algorithm cost two payload sized host transfers per message. An agent skill, migrate-node-to-rosidl-buffer, audits allocations, serialization and stream ownership, and Nsight Systems verification then confirms that no payload sized host to device transfer occurs and that msg->data.get_backend_type() reports cuda. On a Jetson, this decides whether the GPU does the work or sits behind a copy, and it pairs with the transport tuning in the QoS guide and the basics in what ROS 2 is.

A mission state that survives a supervisor restart

Dennis Lanov published an experimental preview of ROS 2 Task Resilience, aimed at what happens to unfinished missions when the process managing them disappears. A ROS independent SQLite core holds mission and step state, dispatch identity and execution attempt accounting, while a ROS facing package adds single host execution ownership, a thin NavigateToPose adapter and a small run, resume and status CLI (ROS Discourse, repository).

The demo is hardware free: a two step mission visits checkpoint A and returns to base, the supervisor exits at a controlled point after submitting the return goal but before recording its result, and a second process reopens the same database and recovers the retained result without resending the goal. The limits are stated plainly: no exactly once guarantee, no cancellation integration, no automatic retries, one mission per database, tested on Jazzy under Ubuntu 24.04 in WSL2 and in simulation only. When the action server’s evidence is gone, the invocation stays UNRESOLVED rather than assuming success. That is the honest version of the problem, and a useful contrast with the goal semantics in topics versus services versus actions.

Describe a robot, get a buildable ROS 2 workspace

RoboStacker is a free web tool that takes a plain description of a robot (base, sensors, computer, tasks) and returns the ROS 2 packages it needs in bring-up order, the pieces still missing such as an odometry source for the EKF, a map for Nav2 or a sensor frame, and a downloadable workspace with the robot model, ros2_control configuration, EKF, Nav2 settings and launch files (ROS Discourse, tool). It covers warehouse robots, skid steer and Ackermann rovers, PX4 and ArduPilot companion computers, and arms in beta, on Humble and Jazzy. Every generated robot is built and tested automatically on each change, and the wheeled ones are driven through Nav2 on simulated motors. No account is needed, and the author wants feedback on whether colcon build works on the generated workspaces. That gap between a package list and a working stack is where integration time goes, the layer the modular stack argument puts value in.

Tactile skins that make human demonstrations transferable

SoTa, from a Stanford group with co-authors at Tsinghua and Cambridge, targets the cost of visuo-tactile data. Dexterous tactile demonstrations need a teleoperated robot, which caps dataset size, and human demonstrations only transfer if the human hand and the robot hand carry sensors that produce corresponding signals. SoTa is a low-cost capacitive skin with full-hand coverage on both, holding a shared layout of 202 taxels across matching finger and palm regions. The multilayer design with fabric electrodes allows in-house fabrication for under 10 dollars in materials per skin, and the sensor keeps over 97 percent of its initial response span after 10,000 loading and unloading cycles, with traces staying continuous through 1,280 tight fist folding cycles (arXiv:2610.02338, project page).

The shared taxel layout is the technical point: it allows human and robot co-training with a common tactile encoder and no learned cross-sensor mapping. Across three contact-rich manipulation tasks, tactile input improved in-distribution success over vision-only policies, and with a fixed robot demonstration budget, adding human demonstrations more than doubled mean success across eight conditions, from 22.8 percent to 45.9 percent, improving all five out-of-distribution conditions. The team plans to open-source the fabrication resources, and the economics here are the argument behind the robot data bottleneck.

Hitachi and Agile Robots co-create physical AI for the factory

Hitachi and Agile Robots signed a co-creation partnership to put physical AI into production. Hitachi contributes its physical AI technology, system integration and edge AI semiconductors. Agile Robots brings robotic arms, mobile robots and the Agile ONE humanoid, running on its AgileCore software for simulation, control and orchestration, and Hitachi will offer the results through its HMAX Industry portfolio (Hitachi, Agile Robots).

The target tasks are the ones conventional automation handles badly: handling parts that differ, loading machines, assembly, changing a line to a new product, moving material between processes. Hitachi signed a separate Fanuc partnership on 30 September with overlapping goals, using Hitachi factories in the Ibaraki region as customer-zero sites and combining Hitachi edge AI silicon with Fanuc robots, with deployment targeted from fiscal 2027. Both deals point at the same limit: imitation learning and digital twins alone are not enough, and the partner that owns the installation owns the data.

An agentic engine for simulation-ready robot data

Awomo-SimDataEngine, from the Awomo-PhysicalRSI Team and collaborators, connects asset and scene generation to demonstration synthesis. Asset services produce rigid and articulated objects with structure grounded parts and joints, scene generation runs two routes (Unravel reconstructs editable scenes from images, SimForge builds environments from text), and a graph native harness coordinates construction, validation and bounded repair, routing a failure back to the module responsible while keeping unaffected scene state. PolicyForge then binds validated worlds to tasks and embodiments to produce replayable demonstrations (arXiv:2610.02274).

The result is the interesting part: co-training with Isaac Sim demonstrations improved a World-Action Model’s overall success on MuJoCo-based LIBERO-Plus from 77.17 percent to 89.43 percent, with goal success up 31.66 points and spatial success up 6.25 points, and more limited gains on long-horizon tasks. Cross-simulator transfer of generated data is the claim to watch, and it is the pipeline layer described in what Isaac Sim does moving a step further from hand-built scenes.

Quick Bites

• NEEDLE, from Juntao Ren, Yifan Hou and Shuran Song, augments robot datasets offline by adding short verified action bridges between recorded observations, using only RGB, proprioception and episode outcomes. On real robots it beat the strongest baseline on each task by an average of 21 percentage points, and the paper is submitted to ICLR 2027 (arXiv:2610.02339, project page).
• DEBIX announced the M8391-01, a credit-card-sized industrial board around the MediaTek Genio 720 with a 9 TOPS NPU, up to 16 GB LPDDR4, a 4 lane MIPI CSI camera input, Gigabit Ethernet with PoE and a minus 40 to 85 degrees Celsius operating range, aimed at edge vision and robotics (CNX Software).
• A ROS Discourse thread asks how commercial or proprietary ROS 2 components are typically distributed, since a proprietary inference runtime, a C API, a ROS 2 adapter and launch files have no standard packaging convention. It has no answers yet (ROS Discourse).