ROS 2 zero-copy, humanoid tender data

The core of ROS 2 is up for open heart surgery, a security vendor found 28 zero-days in the tooling robots are trained and tested with, and a new dataset replaces list prices with signed purchase orders. Also: two small models that write and grade robot behaviour.

REP 0157 would rebuild rosidl for zero-copy

ROS 2 holds a dedicated meeting on REP 0157 on Wednesday, 4 November at 17:30 UTC, after the PMC discussed the proposal on 27 October (ROS Discourse). The draft is titled “Minimal Overhead Messaging Support Using Runtime Agnostic Memory Layouts”, authored by Michel Hidalgo with Miguel Company of eProsima as PMC sponsor, and its status is still Draft (REP 0157).

The diagnosis is blunt. rosidl dates to 2013 and gives each language its own member based representation, so zero-copy cannot cross a language boundary, and variable length members at nonlocal addresses make same language zero-copy so narrow that std_msgs/msg/Header does not qualify. Most robots are single host systems anyway, where copy, convert and serialize costs pile up.

The proposal makes messages language specific views over language agnostic contiguous memory. Storage wraps each member in a rosidl_memory_t carrying an address and attributes, so one layout can describe shared memory, device memory or heap, and shapes carry sizes for variable length members so publishers can be bounded on write. Reference implementations use XCDRv1 as the in-memory representation because it is binary compatible with the wire format most Tier 1 RMW implementations already use.

Compatibility is where this gets expensive. C++ and Python can hide the change behind inline namespaces and properties, but C cannot, and the draft says the redesign necessarily breaks C compatibility while trying to preserve over-the-wire behaviour. Watch the latency benchmarks from the Ekumen-OS reference implementations, not the block diagrams.

VicOne reports 28 zero-days in the physical AI stack

Security firm VicOne says it found 28 zero-day vulnerabilities in less than two days of AI-assisted research across software and tools used to build, simulate, test and operate robots, including projects maintained by Google, NVIDIA and Meta (The Robot Report).

Three quarters of the findings sit in development, training, simulation or validation tooling, so most of the exposure arrives before deployment. The scenarios are the uncomfortable ones: a poisoned pretrained model that behaves in testing and misbehaves on one object later, a tampered simulation scene that lets an unsafe policy pass, and compromised operator software that edits motion planning files. VicOne also demonstrated a vulnerability that stops a robot’s remote-stop command from working.

Treat this as vendor research. The article is sponsored content, the count is described as initial, and some findings have no CVE yet. The engineering point survives: if a compromised component can invalidate a safety test, the test is only as trustworthy as the toolchain behind it. Pin and hash model and scene dependencies, and test whether safeguards still work when the software under them is compromised, which is the argument for a deterministic gate between planner and wheels.

An open catalog of what labs actually pay for humanoids

HumanoidStore published an open dataset of 120 humanoid robots plus a record of 55 public tenders since January 2024, 37 of them awarded across 14 countries, under CC BY 4.0 (ROS Discourse, data, tenders).

The catalog carries sales status, list price with the month it was checked, delivered prices to the US and Germany, degrees of freedom, payload, hands and seller count. The tender table is the more interesting half, because universities and institutes have to publish what they paid. CNRS in Nancy paid EUR 157,700 before VAT for three Unitree G1 EDU units, about EUR 52,600 each, against four bids. The University of Boras paid SEK 1,030,502 before VAT for a single Unitree H2 with one bid. Twelve of 29 competitive awards drew a single bid, the median time from call to award was 122 days, and dealers won 29 of the 37 awards.

That is a planning number rather than a price list: four months from call to hardware, and a real chance of one bidder. Coverage is the EU through TED plus USAspending.gov and the UK Contracts Finder, and models are named only where a notice names them. The publisher earns from seller links and from running purchases for buyers, so read the rows as sourced records, not an audit. For the demand side: North American orders rose 24.8 percent last quarter.

CaPtain post-trains a 4B VLM to write robot code

Researchers at CUHK MMLab released CaPtain, an open pipeline that turns Qwen3.5-4B into a code-as-policy agent with no VLA, no frontier model and no privileged state at test time (arXiv, project page).

The recipe is the one used for coding agents, transplanted to manipulation. Supervised fine-tuning on synthetic trajectories generated with GPT-6 Astra gives a cold start, then multi-turn reinforcement learning teaches the model to chain perception and motion calls. A physical critic reads privileged simulator state the policy never sees and credits each request by how much it changes the predicted chance of success, solving credit assignment across dozens of calls whose reward arrives only at the end.

The reported numbers are specific. CaPtain-4B goes from 0 percent zero-shot to 55.4 percent on held-out LIBERO after SFT and 73.8 percent after RL. It reaches 78.5 percent on LIBERO-PRO, which the authors say beats several frontier-model agents, holds gains of 17.6 to 21.2 points across three perturbation suites, hits 83.3 percent on SimplerEnv WidowX without target-domain training, and runs on a real arm after simulation-only adaptation. Code, data and weights are published.

The appeal is inspectability: a failure is a program you can read, and 4B parameters fit local hardware. The open question is perception noise, since the evaluation is dominated by tabletop suites, which makes this a third data point in the VLA versus alternative architecture question alongside world action models.

Nua ships a 4B model that grades robot tasks

Nua, a Y Combinator backed startup, released Artificer, a 4B parameter decision model that estimates from video whether a robot task succeeded and how far along it is (Nua).

The use case is the unglamorous bottleneck in every robot learning loop. An arm can grasp the right object and drop it in the wrong bin, or stop halfway, and checking by hand does not scale. Artificer is positioned as an automatic scorer for training, evaluation and deployment, where an estimate could tell a controller whether to continue, retry, or move on.

Nua reports results on RoboReward, using source-averaged mean absolute error against reference scores, and on ArmNet268 for success detection with balanced accuracy at a fixed threshold. It claims frontier-level accuracy at 4B parameters, ahead of Opus 5.5, GPT-6.1 Sol and GPT-6 Astra. The post does not print per-dataset numbers, and the model is access gated rather than open weights, so the comparison stays a vendor claim until independent numbers appear.

The category still matters, because reward models are becoming the measuring instruments for robot learning. Cheap reliable scoring means more RL, better data filtering, and regressions caught that would pass a fixed test set. That connects to the data pipeline work most teams already do, where the missing piece is usually a trustworthy label per episode rather than more episodes.

Quick Bites

• Trucks VC has fielded pitches from 81 new autonomous vehicle startups in 2026, about two thirds more than last year, and says foundation models let teams deploy in months on under $1 million, with Bedrock Robotics raising $600 million at a $4.1 billion valuation (TechCrunch).
• Percepto launched Axion, a hardware-agnostic inspection platform that ingests data from third-party drones and sensors and claims more than 160 AI models, part of a pivot to a software-first business under new CEO Julien Bellanger (The Robot Report).
• batchplan 0.5 adds tool paths, pick-and-place, blended motions and live replanning, with batchplan-ros support for Cartesian paths on ROS 2 Lyrical and Jazzy and a pixi run demo Panda example (GitHub, ROS Discourse).
• The DC data collection framework for ROS 2 now runs on Jazzy, Lyrical and Rolling, validates records against a JSON schema, uploads resumably to object storage, and moved from Fluent Bit to Vector (ROS Discourse).
• NVIDIA published Omniverse workflows where an agent builds simulation applications from natural language, including a warehouse model rebuilt from panoramic scans and a gas turbine twin wired to Cantera and SciPy (NVIDIA).