This Week

Things that just surfaced and nobody has checked yet. What lives here is a list of leads and to-dos, not conclusions — so every item states what is being checked, how we intend to check it, and when it will land.

Why this is not a news feed

Reposting the news would make this one more aggregation site, and it would trade the weakest evidence for the fastest update rate. That trade is not worth making. A rumour is only a lead; it becomes an entry after someone checks it. When it cannot be checked yet, it stays on this page with a note about what is blocking it.

Week 40, 2026

LeRobot Humanoid · waiting for the first independent builds
The release hands over hardware, runtime, identification and training environments in one go, but public build records are still thin. "Does it walk once assembled" currently rests on the publisher's word alone.
Checking: real assembly hours, standing success rate after identification, and where printed parts fail under load.
How: watch repository issues and community build logs; promote to an entry once three independent builds exist.
Expected: weeks 42-44.
Berkeley Humanoid Lite · is the CNC dependency real?
A commonly repeated point is that it needs machined parts and is therefore harder than fully printed designs. Accounts disagree on the extent: some say only a few load-bearing parts, others that the joints cannot avoid it.
Checking: which items in the BOM genuinely require machining, and whether printed equivalents with metal inserts cover them.
How: walk the BOM and assembly notes item by item and classify.
Expected: week 41.
Robstride motors · what continuous torque survives real cooling?
The datasheet gives peak and continuous figures, but continuous torque depends heavily on cooling conditions, and a humanoid's joint cavity has almost no airflow. This is the most consistently underestimated item in selection.
Checking: temperature rise and derating thresholds for the same model inside an enclosed cavity.
How: wait for published thermal data; if none appears, record "no public data" rather than guessing on the motor's behalf.
Expected: ongoing.
Sourccey home robot · does the opening match the announcement?
An open-source home robot whose hardware, electronics, software designs and printable STLs were all promised for publication. There is a gap between a promised open-sourcing date and the actual release, and that gap is exactly what these projects should be checked for.
Checking: which files are actually in the published tree, whether editable sources are included, and the licence terms.
How: go by the repository and the official release page, not the marketing page.
Expected: one to two weeks after release.
Chinese VLA models · the gap in cross-language coverage
Several open embodied foundation models have appeared on the Chinese side recently, while English-language evaluation and reproduction records are almost absent — and the reverse also holds. The two evaluation cultures barely intersect, which is precisely where a bilingual site can add something.
Checking: whether weights and training code are genuinely released, and whether reported metrics have independent reproduction.
How: verify licence and file manifest first, then look for independent reproduction.
Expected: entries through weeks 42-45.

How last week closed out

Results only, no process.

  • Promoted to entries: three blueprints (LeRobot Humanoid, SO-101, Thor), three software entries and three hardware entries — each through evidence grading, each stating what we checked and what we did not.
  • Demoted to object-page fields: several projects with content on one side only turned out too thin to stand as entries once checked; the facts we did confirm went onto their object pages.
  • Still open: the machining dependency on Berkeley Humanoid Lite, and continuous torque on the Robstride motors. Both lack first-hand data, so neither becomes an entry yet.
Have a lead?

This page is one of the ways in. The most useful kind of lead is not "so-and-so released something new" but "what this claim says and what actually shipped do not match" — send it with where you saw the original wording, and we will check it.