Fleet APIs, in-orbit assembly and zero-copy rosidl

Three different layers of robotics moved this week. Serve turned its sidewalk fleet into an API, Caltech proved a robot can build a structure in orbit without astronauts, and the ROS 2 community set a date to settle zero-copy across process boundaries. Around them, a five year AI review named embodied intelligence as an open grand challenge, a data vendor productized egocentric video, and ABB and NVIDIA packaged physical AI training.

Serve lets anyone dispatch its sidewalk robots

Serve Robotics launched an AV Delivery Developer Portal that exposes its delivery fleet through a public API, plus a sandbox for testing integrations against simulated robots (The Robot Report, developer docs). The company says it has deployed more than 2,000 robots across the US and Australia, reaching about 3 million people and serving more than 4,000 restaurants.

The integration is three calls: request a quote for a pickup and dropoff, create a job from that quote, then poll until the job reaches dropoff_complete. The published details are more interesting than the pitch. A quote expires after five minutes and works once, so reuse returns 409 used_quote. Retries need an Idempotency-Key header. The rate limit is 600 requests per minute. Money is integer cents and coordinates are WGS84 decimal degrees.

The sandbox is the unlock: it runs the full job lifecycle against simulated fleets, with endpoints to spawn robots, teleport them and advance a job stage by hand. You can build a complete integration before a commercial agreement exists, then swap the base URL and key. Live dispatch still needs a service area, a production key and fleet access. If you build on top, dispatch becomes someone else’s problem, the same tradeoff as in how Nav2 handles navigation one layer down.

Caltech assembles a truss in orbit

A Caltech robot has assembled part of a structure in space without astronauts, which the project lead calls the first autonomous construction of its kind (ASSEMBLY, DARPA NOM4D). The ASTRA payload launched on 30 March 2026 aboard Momentus Vigoride 7 on SpaceX Transporter-16, then spent six months working toward a 1.4 meter diameter reflector antenna from composite tubes, aluminum joints and perforated Kapton film.

An autonomous manipulator pulls struts and joints from motorized dispensers and repeats identical bays on an assembly plate. A full truss needs 11 bays. Time ran short, so the team built only the first section. Ground control had seven or eight communication windows of about 15 minutes a day, checked results against downlinked photos, and rehearsed fixes on a flight spare before sending workarounds up.

The problem generalizes. Work too long or variable for one open loop plan needs verification at each step, which is also the lesson in certified stops and unverified inputs. DARPA’s NOM4D program started in 2022 aiming at structures hundreds of meters to kilometers across, beyond any rocket fairing. Watch the December mission end for what the team simplifies, the manipulator or the part feeding.

REP 0157 puts zero-copy inside rosidl

Open Robotics scheduled a follow-up meeting on REP 0157 for 16 November at 17:30 UTC, after a 4 November session left open questions (ROS Discourse, REP draft). The proposal redesigns rosidl so ROS 2 gets native zero-copy transport across process boundaries, with language agnostic memory layouts, strict message constraints, XCDR as the in-memory representation, backward compatibility, and support for C, C++ and Python.

Three signals came out of the first meeting. Member based message access versus method based access is heading toward supporting both, then deprecating the member form. Serialization may be decoupled from middleware, with a Zenoh prototype planned to test a non-DDS path. Benchmarks showed zero-copy gains become negligible for small payloads, which matters because most robot topics are small.

Miguel Company then proposed first class views, built over a message or a binary blob. The next agenda covers that, migration to method based access, a generic type support context with payload bounds, and open questions on buffer integration and loaned messages. It is the cross-process version of the same idea in ROS 2 intra-process communication and composition, with a stricter contract.

AI100 puts embodied AI on the grand challenge list

The third AI100 report, “AI Goes Mainstream: The Scaling Era and the Rise of Generative AI”, names physical AI and embodied intelligence among six inspiring open grand challenges (AI100, grand challenges section). The panel credits vision-language-action models and simulation for narrowing part of the sim-to-real gap since 2021, then lists what remains: long horizon planning under uncertainty, dexterous manipulation of diverse objects, and formal performance guarantees.

The numbers are worth keeping. China deploys more industrial robots than the rest of the world combined and makes about 80% of the world’s humanoid robots, a segment the report puts at 2% of the market. More than half a million factory robots are installed each year, in constrained environments. On the BEHAVIOR-1K simulated household benchmark, no system scores above 26%.

The useful framing is the gap between demonstration and deployment. A policy that works in a demo has not answered what happens when the object is wet or the task runs for 20 minutes. That is the same argument made in the robot data refinery and in world-action models and their bottleneck.

Memories.ai sells the pipeline, not just the video

Memories.ai is selling robotics teams a standardized egocentric video pipeline: capture hardware, a uniform collection process, automatic annotation and a data lake aimed at VLA models, world models and humanoids (The Robot Report). The San Francisco company gives collectors a modular camera setup, then labels footage with its own models, including Memories-H0 for hand pose and EgoAction for action segments, and says it has published at ICLR, NeurIPS and EMNLP.

The NVIDIA angle is worth watching. Memories.ai says it uses Cosmos to interpret incoming video, curates that data, and post-trains Cosmos for customer applications. Treat it as a vendor statement for now. Hand pose from a head mounted camera is not end effector pose on a specific gripper, and retargeting across embodiments is unsolved. Ask how label accuracy is measured and how much of a dataset survives the embodiment gap.

ABB and NVIDIA build a physical AI curriculum

ABB Robotics and NVIDIA are developing an education program for engineers and manufacturing workers, mixing theory with hands-on work in RobotStudio and Omniverse, with a certificate at the end (ASSEMBLY). A white paper released with it reports that 84% of companies have not redesigned jobs around AI capabilities, while 53% are working on AI fluency.

ABB also said RobotStudio HyperReality, built with Enodo Robotics on OpenUSD, is now available to 60,000 RobotStudio users and claims 99% accuracy in closing the sim-to-real gap. That is a vendor claim, and the definition of accuracy matters more than the number.

The developer point is the tooling default. If workcells are designed and validated in simulation, the skills that matter are scene construction, parameter tuning and knowing which physics settings change the answer, which is a different job from online programming at the cell.

Quick Bites

• Stingray Marine Solutions has more than 3,500 laser armed units in Norwegian salmon pens, protecting over 112 million fish and covering 20 to 25% of the country’s farmed salmon, up about 1,000 in a year (Tom’s Hardware).
• Jazzy Jalisco sync is planned for 13 November with 752 packages ready and no regressions on Ubuntu Noble (ROS Discourse).
• Humble sync follows on 16 November with 416 packages and 10 open regressions, pending a proxsuite fix (ROS Discourse).
• NVIDIA donated OpenShell, a runtime for securing autonomous agents, to the CNCF Sandbox and is contributing to Google’s Agent Substrate (NVIDIA).