Mobile robot safety rules, open hand

Today: the safety rulebook for industrial mobile robots gets its user-facing chapter, an open 15 degree of freedom hand gains a ROS 2 driver, and a benchmark report argues that one success number tells you little about what a model can do.

A3 closes the user-side chapter of AMR safety

The Association for Advancing Automation has released the final part of its industrial mobile robot safety standard. ANSI/A3 R15.08-3-2026, Part 3: Use of IMR Applications, covers what happens after a mobile robot is deployed and is aimed at fleet operators rather than robot builders. The requirements include risk assessment, management of change, employee training, and periodic reviews and inspections, so acceptable risk is maintained through day-to-day operation and equipment modifications (ASSEMBLY Magazine, Automation World).

Part 1 covers the robot itself and Part 2 covers IMR systems and applications, so Part 3 completes the lifecycle view. It was developed alongside Part 3 of the R15.06 industrial robot safety standard, giving a site running both fixed arms and mobile robots consistent guidance (A3 standards index). For integrators the consequence is process rather than code: a fleet that adds a payload or reworks a workcell now sits inside an explicit change-management and inspection loop.

An open hand you can print and drive from ROS 2

Vittorio Lumare posted the Yeah Hand (also called Rebelia V2) to ROS Discourse: an open source, tendon-driven robotic hand with a ROS 2 driver. It has 15 degrees of freedom driven by five actuators (Waveshare ST3215HS bus servos), torque sensing on each finger for adaptive grasping, 3D printed TPU fingers and palm, and ESP32 firmware reachable over Bluetooth serial. Mounting flanges cover ISO 9409-1-50-4-M6 for UR, KUKA, FANUC, Techman and Yaskawa arms, plus Agilex and PCD-40 patterns (ROS Discourse).

The licensing is deliberately copyleft: CERN-OHL-S-2.0 for hardware, GPL-3.0 for software, CC-BY-SA for documentation, which the README describes as a defensive publication against third-party patent claims. The project states a build cost target below 2,000 EUR and succeeds the OSHWA-certified Rebelia V1 (repository). The ROS 2 package is tested on Jazzy and exposes per-servo torque limits at runtime (driver repository).

This is still early work: a handful of commits, single-digit stars, and V1 remains the only OSHWA-certified version. It is a useful counterweight to the industrial argument all the same. Boston Dynamics took Atlas to four fingers and 13 directly actuated degrees of freedom and spent the design budget on ruggedness and manufacturability. A 15-DOF tendon-driven hand under 2,000 EUR attacks the same problem from the other end, and the ROS 2 primer covers the interfaces such a driver exposes.

Tesla explains the Robotaxi night limit

Late on Friday, Tesla’s Robotaxi account said the Austin service now runs until 11pm, an hour later than before. Elon Musk explained what holds it back beyond that: “The main thing we’re trying to solve is making sure that we don’t run over pets when they’re hard to see at night. Literally trying to avoid grey kittens on grey tarmac in the dark.” That is Electrek’s account, quoting Musk’s post on X (Electrek).

The engineering is not new, and it is worth restating for anyone building perception. A camera is passive: it needs light and contrast to separate an object from its background, so a low-reflectance animal on wet asphalt at night is close to the worst case. Lidar fires its own pulses and times the returns, giving range and 3D shape independent of scene brightness, while radar adds velocity and survives fog, dust and glare. Electrek reports that Waymo’s sixth-generation system carries 13 cameras, 4 lidars and 6 radars, and that Tesla removed radar in 2021 and has never shipped lidar. For robotics the takeaway is the sensor-fusion argument mobile robot teams already live with: an AMR in a dim warehouse or a drone at dusk hits the same failure mode, and the fix is a second modality that does not depend on reflected visible light.

Benchmark means hide which tasks a model fails

Fig published a technical report, with authors from Fig and Georgia Tech, that scores frontier models task by task instead of benchmark by benchmark. Seven models were run on 177 VisualWebArena computer-use tasks, and three across four physical domains: VLABench for manipulation, Assembly101 for procedure understanding, IndEgo for next-action anticipation and Bench2Drive-VL for driving (Fig).

The finding is structural. Across all 21 model pairs in the web sweep, the lower-scoring model solves at least three tasks the higher-scoring one fails, and in the widest pair it solves 21. Astra leads at 139 of 177 (78.5 percent), with Fable 5 one task behind and inside the report’s plus or minus three point noise band. One reported model update moved the mean by 1.1 points while 36 episodes flipped. The same unevenness shows physically: models that beat Gemini 3.5-Flash on every domain average still lose some task categories. The authors released RIDGE, item-level results and model traces for all five domains.

A number like 78 percent summarises a distribution containing tasks the model fails and tasks it alone solves. If you are choosing a vision-language-action model for a pick-and-place cell, jagged competence decides the deployment, and the data bottleneck is why your own task distribution, not a leaderboard, ends up being the real test set.

Destro sells the orchestration layer, not the robot

Destro emerged from stealth with an 8 million dollar seed round led by Base10 Partners and Bonfire Ventures, with CoFound Partners also participating. The company does not build robots. It builds an intelligence layer that directs the whole operation, including the humans: a Vision operating system based on open-weight vision-language-action models that drives the machines, and a Mothership layer that assigns work to people, carts and trucks alongside them (TechCrunch).

At a Yusen Logistics facility in the Pacific Northwest, Destro started with three cart-moving robots from Miva Robotics, directing where full loads go while workers unload trucks into mixed carts for cross-docking. Richard Brunelle, who runs automation across roughly 30 US sites for Yusen, told TechCrunch that two better-known robot startups could not fit the workflow, one because its robots could not handle loading and unloading, the other because its fleet tools still needed a person to orchestrate everything. Destro is expanding to 26 robots there and starting a second pilot with 17 robots in Southern California.

Those are Destro’s numbers and its own site is still a placeholder, so TechCrunch is the only account. The architectural point is the one developers should take: it is the layer the modular robot stack argument describes, where value accrues to the harness that assigns tasks across robot bodies and the people beside them. The open question is whether foundation model and hardware vendors sell into that harness or absorb it.

Quick Bites

• RobotCAD v12.10.0 adds a RobotCAD MCP server for agentic robot assembly, part of a release note covering many recent versions of the ROS-adjacent CAD tool (ROS Discourse).
• A ROS Discourse thread asks how to validate a hand-eye calibration beyond reprojection error, noting that MoveIt Calibration, easy_handeye and Kalibr print a transform or an error term without a verdict, with no accepted answer yet (ROS Discourse).
• IEEE Spectrum’s Video Friday features a bioinspired albatross robot that falls, spins, self-rights and sails away, alongside a modular arm (IEEE Spectrum).
• Samtec introduced connectors for the SGeT oHFM.c standard at 0.635 mm pitch, targeting 64 or 112 Gbps PAM4 for FPGA and system-on-module interfaces (Design World), the quiet prerequisite for feeding high resolution cameras and lidar into onboard inference.