Files: MD · PDF

Landscape doc vs How Humanoids Learn - coverage audit

Notes from a chat with Claude, October 3–4, 2026.

Question: How much of the information and data in "Open-Source Humanoid Robotics Landscape (Oct 2026)" is in, or missing from, the site "How Humanoids Learn"?


About a fifth of the landscape doc is on the site in some form, and only a tenth of it fully. Roughly 80% is missing. The site covers the well-known core: GR00T, openpi/π0, Isaac Lab, MuJoCo, Newton and the main datasets. The open-hardware catalog and the project-tracking data are almost entirely absent.

The numbers (287 projects, 1,882 table cells, 547 prose sentences; the methodology section is excluded)

  • Projects: 110 (38%) are named on the site, and 101 of those have at least one fact there. The other 177 (62%) never appear.
  • Table data: 11% is fully on the site, 5% partly, under 1% disagrees and 83% is missing.
    • Descriptions do best: 30% are at least partly there.
    • Specs and numbers: 14%. Licenses: 10%.
    • Stars, status and last-release dates: 3%.
  • Prose: 5% fully present, 27% partly, 2% disagrees, 67% missing.
  • If a partial match counts as half, the site holds about 16% of the doc's information.

By section (share of facts at least partly on the site, and projects named)

SectionFactsProjects
Simulators54%13 of 18
Foundation models37%14 of 24
Datasets & benchmarks36%23 of 44
Middleware33%5 of 13
Whole-body control30%7 of 20
Grippers, touch sensors, gloves22%8 of 21
Stalled projects21%21 of 53
Teleop systems17%4 of 10
Closed-hardware humanoids16%5 of 13
Activity scoreboard12%51 of 161
Actuators10%1 of 7
Arms8%5 of 16
Dexterous hands3%4 of 14
Open humanoid hardware3%1 of 14

The Summary (43%) and the trade-offs and recommended stacks (69%) are mostly partial matches. The site makes similar points without the doc's specifics.

Biggest gaps

  • Open humanoid hardware: ToddlerBot, Berkeley Humanoid Lite, Asimov, Roboto Origin, HopeJR, OpenLoong and iCub. The height, weight, DoF and price table is entirely absent.
  • Hands and hand data: LEAP, ORCA, RUKA, Aero Hand and AmazingHand; DexUMI, DOGlove, AirExo-2, AnySkin and 9DTact.
  • Arms and actuators: there are no arm specs at all; the only arm figure is one OpenArm price. reBot, PAROL6, AR4, Koch, moteus, SimpleFOC, ODrive and OpenQDD are missing.
  • 2026 open models, mostly Chinese: X-VLA, RDT2, MolmoAct2, Xiaomi-Robotics-0, GigaBrain, LingBot, InternVLA, EO-1 and OpenDM. Only 1 of 44 LIBERO, SimplerEnv and download figures is there.
  • Open whole-body-control tools: mjlab, holosoma, ProtoMotions, MimicKit, PBHC, VideoMimic, Pinocchio and Crocoddyl.
  • Tracking data: stars, status and last release for all 161 projects, and license terms for most of them. Only Isaac Gym is called legacy, out of 22 stalled project groups.

Where they disagree. There are about 14 distinct points; these five matter:

  1. GR00T N1.5/N1.6 weights: the doc says they were non-commercial until N1.7 in April 2026. The site says N1–N1.6 already allowed commercial use.
  2. GMR: the doc calls it kinematic retargeting. The site groups it with methods that enforce contact and dynamics, which looks like a mistake on the site.
  3. SONIC on a real G1: the doc says 100% on 50 trajectories and 19 of 20 on apple-to-plate. The site says 99.2% and 90%.
  4. Isaac Lab 3.0 early access: the doc says Sep 16, 2026. The site says March 2026, with betas through June.
  5. RLDS, HOVER and ASAP: the doc lists them as legacy or dormant, but the site still recommends them.

The rest are minor: LeRobotDataset v3.0 (Sep vs Oct 2025), Octo (2023 vs 2024), Galaxea Open-World (150 vs 227 tasks), SAPIEN's maintainer (Hillbot vs UCSD), and the G1 price ($16k at the 2024 launch vs from $13.5k now).

Most of the gap is catalog detail rather than missing ideas: the site explains how humanoids learn, while the doc tracks projects.

To get these numbers, I matched every project name against the site. Sub-agents then checked each fact about the 110 named projects against the site's text, players atlas, timeline and widgets, and I spot-checked their calls.

Sources: