Executive Summary

The moment the list went up on LinkedIn can be pinned down by arithmetic: 16:31 UTC on 31 October 2026. LinkedIn attaches an activity identifier to every post, and the top 41 bits of that number are a millisecond timestamp. The headline read "Which robot action models are best at learning new tasks?", and the author wrote that he had reviewed thirty robot action models and narrowed them to fifteen. Seventy-two reactions and two comments followed. It is one person's reading, and this report is not aimed at that person. What it is aimed at is why lists like this get made so quickly and travel so fast.

One thing deserves to be said up front, in fairness. The list does not hide its reasoning. The author named four things he weighed when putting the models in order: whether a model handles a new task zero-shot, whether it learns from a handful of demonstrations, whether it adapts on site, and whether it transfers to a different body. He even listed where to look next. What is missing is the scores on those four axes and the procedure that produced them. Across fifteen lines there is not a single number.

We opened embodied.guide too, the site named as further reading. It does not publish rankings. Its own description says "Independent by design — No paid listings. No affiliate links. No manufacturer rankings." It is a catalog of specifications for 366 pieces of robot hardware (255 from China, 34 from the United States, 13 from India), which is a different axis from a model ranking. So there is, in practice, no outside basis holding up the order of the list.

The nationalities the list assigns are also hard to carry over as given. The post splits the fifteen into nine American and six Chinese, yet Gemini Robotics 2 in third place was built by Google DeepMind, which is headquartered in London. The count changes with the counting rule. This report does not use the regional tally as a load-bearing part of its argument.

1.1"Roughly Two per Week" Turns Out Not to Be an Exaggeration

The post contains the sentence "New robot action models are appearing at a remarkable pace, roughly two per week". That is the author's estimate, not a tally. But checking the first announcement date of each of the fifteen models on the list, one by one, the estimate does not err on the high side. All fifteen arrived between April and October 2026, inside seven months. Ten of them are packed into the three months from August to October alone.

We opened each company announcement, paper and model card as a primary source, confirmed the release date, and counted by month. This count is true only within the fifteen the author chose. It cannot tell you how many models the field as a whole ships per month. Where one entry covers two versions (Dyna-2 and Dyna-2.1) we counted the earlier release date.

On the academic side the pace is tighter still. Querying the arXiv search server restricted to the cs.RO category and counting papers whose abstract contains the phrase "vision-language-action" by month, the figure rises from 74 in April 2026 to 276 in September. The phrase "world action model" went from 8 to 90 over the same stretch. This is a crude count that only checks whether a phrase appears, so surveys and passing mentions are mixed in, and its scope is not the scope of the monthly bars above. These are numbers you must not subtract from or divide by one another. Read for direction only, they say that product announcements are pouring out of a field that has not yet settled on what to call itself.

That second phrase, "world action model", points to a family that first sketches what will happen next and then picks its motion to fit. WAM is the short form, and several models on the list describe themselves that way. A survey posted in September 2026 (arXiv:2609.16074, sixteen authors) draws the boundary between VLA and WAM and separately catalogs the datasets, benchmarks, metrics and protocols used in evaluation. We opened only the abstract of that paper, so the details of the taxonomy are better read at the source. What can be read here is that an effort to organize the measuring side already exists, even while the naming is still unsettled.

This pace is what creates demand for lists. When Pebblous stood in the same place seven months ago and compared how three teams designed a robot brain, there were three things to compare. Today one public board alone holds seventy-nine rows. Comparing three means taking the architectures apart; putting thirty in order requires a shared standard. Which makes the next question this one: was there really no yardstick?

One day before the list went up, a shared yardstick built to answer the same question had already finished an update. It is called RoboDojo, and its paper landed on arXiv on 5 July 2026 (arXiv:2607.04434) with forty-four authors. It is built from 42 simulation tasks and 18 real-robot tasks, and at launch the researchers integrated and ran thirty policies themselves to fill in the scoreboard. The abstract says, verbatim, "We integrate 30 policies into XPolicyLab and evaluate them on RoboDojo, establishing a public leaderboard". The snapshot date stamped on the board is 30 October 2026, one day ahead of the LinkedIn list. As of that date, 66 sim rows and 13 real-robot rows — seventy-nine in all — were sitting on one shared procedure.

What it measures is spelled out too. The paper names five capabilities: generalization, memory, precision, long-horizon task execution, and open vocabulary, meaning the model understands objects the instruction never named. The board's table has six columns, because it splits generalization into a fixed setting and a perturbed one. Both numbers are correct. Mixing them makes one of them wrong, so this report works from the paper's five and writes six columns only when pointing at the table.

On the real-robot track the columns mean something else entirely. They are not capability axes but three robot bodies (ARX X5, Piper, PiperX): the same tasks put on different arms. So a sim score and a real-robot score must not be placed side by side and subtracted. The tasks differ and the bodies differ. Where two such numbers appear in one paragraph of this report, there is no subtraction between them.

RoboDojo was not even the only place standing open. There is a second public board, RoboTwin 2.0, which has no real-robot track, runs in simulation only, and is built from 50 tasks. The update date recorded in its board data is 30 October 2026, the same day as the RoboDojo snapshot. The table below puts the two of them — three tracks in all — in one place and compares row counts, what is being measured, and whether the conditions are written next to the scores. The last column of that table is the question this report keeps asking to the end.

We pulled the public data of both boards ourselves on 2 November 2026 and counted. RoboDojo carries every row inside the JavaScript bundle that draws the page; RoboTwin lets you download robotwin_leaderboard.json as is. "Not in the board's table" in the rightmost column does not mean there are no conditions. It means you cannot read them from where you read the scores. We did not get as far as the submission and evaluation section in the body of the RoboDojo paper, so the allowed scope of fine-tuning and the number of attempts remain unconfirmed.

2.1The Board That Wrote Its Protocol into the Data File

RoboTwin 2.0 sits alone in a different cell of the table above. Open its leaderboard data file and the conditions are written right next to the scores: the robot fixed to a single Aloha-AgileX, training data of 2,500 episodes across 50 tasks at 50 each, and evaluation run at 100 attempts per task in both a clean environment and a perturbed one. What matters is that the place showing the scores and the place recording the conditions are the same place. Conditions existing somewhere and conditions sitting next to the number are two different things.

There is one more thing. RoboTwin stamps the name RoboTwin Team on the baseline rows its own team ran: 16 of the 33 rows. RoboDojo leaves the same fact as a blank. One board records the distinction and the other does not, and section 4 is the story of that blank. The argument that failure only becomes visible when baselines sit in the scoreboard is one Pebblous also took up in the fusion reactor design dataset BoSS. In that case, putting a do-nothing predictor in the table was what revealed that two of three models lost to it.

2.2Who Runs It, and What Has Not Happened Yet

RoboDojo states who operates it and how it is funded on its front page. Verbatim:

The word that catches on reading is will. Co-governance is not the present state but a plan written down as intent. So rendering this as "several institutions jointly operate it" goes one step further than the source. What is confirmed today is that one non-profit foundation maintains it and that it has declared it will take no commercial sponsorship. The same page also carries a sentence saying that official listing requires a standard verification procedure.

The real-robot side has a structure attached to it called RoboDojo-RealEval. According to the paper, you connect remotely to the cloud, run on standard hardware, the scene is reset to the same state every time, and the deployment interface is unified. The most expensive part of real-robot evaluation is recreating identical conditions, and that is what has been moved onto shared infrastructure. On that infrastructure, as of 30 October 2026, the top real-robot entry is Shanghai AI Lab's InternW0-Δ at a 31.70% success rate. That is the highest number among the thirteen rows.

Pebblous has previously covered what happened to the ranking of twenty-two models when two pairs of physical AI benchmarks were merged. That report was about benchmarks overlapping one another. Here the trouble is the reverse. Overlap is not the problem; who turns up on a benchmark that stands perfectly open is. Same ground, different question.

We put the fifteen lines of the list and the seventy-nine rows of the board side by side and matched the names. The result fits in one sentence. The top seven of the ranking hold no row on either board. No submitted row, and no baseline row run on their behalf by the operating team. That is what the entirely empty upper seven lines of the table below look like.

The percentages are the board's success-rate column, from the 30 October 2026 snapshot. ⚠️ marks a row whose name does not line up. There is no basis for treating the board's "GO-3 Nano Preview" as the same thing as the list's "GO-3 Preview", and the same goes for "WALL-OSS-1.0-lite" and the unversioned "LingBot-VLA". So we did not carry the scores on those three rows over as scores for the list entries. The only entries whose versions can be said to match exactly are GR00T N1.7, VPP2 and Galaxea G0.5 — three of them.

The count runs like this. Six of the fifteen have a name on the sim board; one has a name on the real-robot board. That one is GR00T N1.7, at a 1.70% success rate. And the overlap is nearly as thin in the other direction. Seven of the board's top ten rows are names that appear nowhere on the list. Two lists claiming to answer the same question barely know each other.

The comparison table above, redrawn. Rows whose names do not line up, such as positions 9 and 10, are colored in as present. Match versions strictly and the number of filled cells drops to three. The one criterion separating empty from filled is whether the name appears on a board.

3.1Being Absent Is Not the Same as Being Unable

Let us block the easiest misreading of this table first. Absence from a board is not evidence that the model is weak. It means only that there is no way to measure it from outside. Nowhere in this report is there evidence that the top seven are bad, and several of them aim at tasks the board does not cover at all. A living room, a logistics warehouse, an hour-long continuous job: none of that is inside the 42 tasks. There are clearly cases where staying off a board is the rational choice.

But that is exactly why what the ranking rests on gets harder to see. With no numbers on a shared yardstick, what remains is the number each company produced on its own tasks, and those numbers cannot be compared with each other. A list is what you get when numbers of that kind are lined up in one column. Pebblous has seen the same structure in other fields: a report on why a coding AI score measured on public repositories cannot be carried over to in-house code, and on the robotics side, a case where the model judging success and failure was itself unsteady. When the measuring instrument wobbles, the ranking built on top of it wobbles more.

Stop here and the conclusion is that the list is thinly grounded and the board is solid. But opening the board all the way showed that conclusion to be only half right.

The board's table has one more column: a field recording who submitted the entry, and filled rows and blank rows are mixed together in it. Not knowing what those blanks are leaves you reading the board as a single ranking, which would mean this report committing the very error it is pointing at. Fortunately the site holds the answer. Inside the bundle that draws the page sits the site's own changelog, 28 entries of it.

The oldest entry is dated 6 July 2026. It opens with "Launch the first sim leaderboard with 30 models" and names all thirty one by one. The twenty-seven entries after it all take the form "Add [model name] (score, contributed by [institution])". Match them up and every one of the 24 blank rows falls inside those thirty launch models. The remaining six later acquired a submitter entry. Twenty-four plus six is thirty.

So rather than "the blanks look like baselines the operating team ran", we can state flatly that the site's own changelog records them as such. And one more line follows from that. GR00T N1.7 is inside those thirty launch models too. Its 2.85 points at 52nd on sim and 1.70% at 9th on real robots are not figures NVIDIA submitted but figures the RoboDojo team ran with its own hands. Put that number on one line with a vendor-submitted row scoring 61.90 and write "GR00T is actually 52nd", and the sentence is false.

4.1The Two Groups Barely Overlap

We split the 66 rows into three groups by the submitter field: 39 rows submitted from outside, 3 rows naming the operating team, and 24 blanks. The three are not evenly mixed across the board. Not one blank row appears in the top thirty-one. The highest-placed blank row is Pi-05 at 32nd (6.93% success rate), and the blank rows run from there down to 66th at the bottom.

The smallest of the three groups is worth setting out on its own. The rows carrying the operating team's name directly in the submitter field are GPT-6-Astra at 18th (22.48% success rate), DeepSeek-Flash at 51st (1.92%) and GPT-5.5 at 56th (0.88%), and all three are general-purpose models run on the tasks rather than robot-specific policies. None of the thirty models run at launch places above that 18th position. The board does not disclose in what configuration these three were operated, so this must not be carried over as a claim that general-purpose models beat specialized policies. In the figure below the three groups are laid out in board order.

The 79 rows of the 30 October 2026 snapshot, colored by the submitter field alone. The three darker cells in the sim strip are the 18th, 51st and 56th positions from the paragraph above, given a different color from the blank-field rows. On the real-robot side the split is sharper still: the three rows with a submitter are 1st, 2nd and 3rd, and the other ten run from 4th down.

4.2Reading This Split as Cause and Effect Gets It Wrong

The temptation here is to jump to "submitting gets you a good score". The public material does not let you write that. At least three causes are tangled together and there is no way to separate how much each contributed. The first is selection: you submit when it looks like it will go well. The second is generation gap. The 24 blank rows contain policies that were already public when the board opened — ACT, RDT, H-RDT, TinyVLA, OpenVLA-OFT, Pi-0, GO-1, SmolVLA — while rows arriving from outside were mostly added between August and October 2026. The third is differing conditions: we could not confirm whether the allowance for fine-tuning and extra data is the same on the submission side as on the baseline side.

Evidence that generation and submission move together sits inside the same board. Galaxea submitted G0.5 itself and placed 21st, while the previous generation G0 sits at 44th as a blank row. Xiaomi likewise has Robotics-1 as a submitted row at 22nd and Robotics-0 as a blank row at 41st. Even within one company the newer version goes up and the older one stays in the baseline slot. So there is exactly one sentence available here. The two groups barely overlap. Why cannot be settled from public material. Which is why this column is not a single ranking.

4.3One More Distinction the Table Does Not Show

The changelog carries a tag that the table does not. Six entries have "(agent track)" written after them, and those six happen to be exactly 1st through 6th on the sim board. PhysicalRSI at 9th is annotated "as baseline solution (Agent)". None of this distinction is visible in the board's table. A row running a single policy and a row running several components stitched together are sorted on one criterion in one table.

So writing "the board's top entry Q0-Preview is absent from the list" is factually correct and still takes the reader somewhere wrong. The accurate sentence is this. The highest row entered as a single policy is WALL-OSS-1.0-lite at 7th with a 34.97% success rate, and that model is 9th on the LinkedIn list. This one line cuts into the impression left by section 3 that the two lists do not overlap at the top. That impression should be cut.

4.4Two Claims of First Place, Both True on the Day

Within a single month, two companies each announced first place on the same board. It looks like a contradiction, but rewinding the board state day by day using the listing dates in the changelog shows both were saying something true.

This timeline is a reconstruction. We rewound it by matching listing dates in the changelog against the scores in the 30 October 2026 snapshot, assuming no row was ever deleted and no score ever re-run. The changelog does contain one score-update record, so the assumption is not airtight. But the highest success rate among the thirty launch models is 6.93%, too low to disturb the October leaders, so the rewind of the upper part is on the safe side.

ROBOTERA put out a press release on 9 October 2026 and attached the qualifier "Without Additional Data or Agent RSI" to the headline. That is a strangely fussy qualifier for promotional copy. But on the board that day, the entry at 2nd directly below VPP2 was PhysicalRSI, the row the changelog annotates "as baseline solution (Agent)". So the qualifier is not decoration but a technical distinction: it climbed above that row without being an agent entry. A vendor press release recorded a distinction the board's table concealed. This was the most unexpected thing we found.

AGIBOT was right about its own day too. At an event in Shanghai on 29 October 2026 it announced that GO-3 Preview held first place on three boards, and for RoboDojo it said first on both task success rate and process score. Rewind the board to that day and both are correct, at 56.42% success and 61.90 points. That first place lasted three days. Q0-Preview went up the next day and pushed it to 2nd, and the LinkedIn list appeared the day after that. Neither party said anything false. With different dates both claims can be true, and both were. The problem is only that a ranking claim without a date leaves the reader unable to reconstruct this.

4.5Even Inside One Table, How Much Gets Recorded Varies

That AGIBOT announcement contains one more small mismatch. The announcement calls the model "GO-3 Preview", but the board row the company itself submitted is named "GO-3 Nano Preview". The party that dropped the version marker is not some third party but the company that built it. The habit of reading the same name as the same thing breaks here.

How much of the conditions gets recorded also varies row by row. Of the 66 rows, exactly one — the DeepSeek-Flash row — has even a single line of evaluation conditions attached, reading "10 episodes / task": ten runs per task. The other sixty-five carry no such note. Inside one table, the conditions are knowable for some rows and unknowable for others.

Open RoboTwin 2.0 and the same shape appears. Of its 33 rows, the top ten are all ten submitted from outside. Three rows appear to overlap with the LinkedIn list, and of those only the top entry, GO-3 Nano Preview (85.01), is a row the vendor submitted; π0.5 (58.33, 16th) and GalaxeaVLA (37.72, 25th) are both baselines run by RoboTwin Team. This must not be written up as "the same pattern confirmed on two independent boards". Both boards run on the same policy integration layer, XPolicyLab. RoboTwin records XPolicyLab standard interface in the evaluation-path field of its board data, and the RoboDojo paper's abstract likewise says that integrating a policy into XPolicyLab once makes it measurable on both the simulation and real-robot sides. What can be said goes as far as this: it is not independent confirmation, but it is not a quirk of one board either.

Of the three places where AGIBOT claimed first, we could not find a public leaderboard for RoboColiseum. A baseline comparison table is published, but there was no way to confirm it is the same table the first-place claim rests on. Two of the three we checked down to the date; one remains unconfirmed. We leave this sentence here so as not to write up an unconfirmed thing as confirmed.

Instead of re-ranking, we count evidence. The fifteen stay exactly where they are, 1st through 15th, and six columns record only what each model has handed to an outsider. Can you get the weights? Is there a technical document? Is there a score on a yardstick someone else built? Are the evaluation conditions written down? Is there any trace of a third party touching it? Can you use it commercially? Not one of the six columns measures how well a model performs. Every one of them measures checkability.

○ present · △ conditional or partial · ✕ not published · — no weights exist, so the question does not arise. Every cell was judged on 2 November 2026 by opening the company announcement, paper, model card and license text directly. A △ in the "Tech doc" column means a blog post and nothing more. We write "not published" rather than "none" for the same reason given in section 3.1: not visible and not existing are different things.

Read the table down its columns and one thing jumps out. For the top seven of the ranking, all seven cells in the weights column are empty, and all seven cells in the shared-score column are empty too. The models whose weights can be obtained in some form are clustered toward the bottom of the list. This is not a route to "so the bottom is better". What the table speaks to is not how well anything performs but whether there is any way to check it from outside.

5.1"Open Weights" Comes in Grades

The ○ and △ in the weights column must not be lumped together. The same word "open" means entirely different things depending on the terms. Five grades are mixed in here: weights you can take, modify and sell; weights you can take but cannot trade on; weights released only in part; weights you can only call at an hourly rate; and weights that have merely been promised for release at the end of the year.

Licenses were judged by reading the model cards and repository texts directly. Unitree UnifoLM-WLA-1.0 carries an Apache 2.0 license but releases only part of what it has, so the weights cell in the main table above is marked △.

GR00T N1.7 occupies the most misread spot in this grade table. The code repository is Apache 2.0, but the model card says something else: "This model is released under the NVIDIA Open Model License Agreement." On top of that, the model's vision-language backbone Cosmos-Reason2-2B is a repository that requires accepting terms to access. Go to collect it thinking it is Apache 2.0 and there is one more door. Did that door hold back reuse? The opposite, as it turns out. Section 6 has the numbers.

5.2It Is Not Companies That Closed, It Is the Frontier

π0.7 at 4th is absent from the board. But π0.5 is on it: 4th on real robots (12.80% success rate), 32nd on sim, and downloaded more than twenty thousand times a month on Hugging Face. The previous generation is open and the current one is closed. That shape repeats across all fifteen. There turn out to be fewer cases than expected where you can say "this company is closed" at the company level. Instead, it is the most advanced version in each lineage that is closed. What closed is not companies but the frontier.

Published scores call for the same caution. Isaac 0.5 at 15th put out a figure of 97.2% on a shared task suite called LIBERO, but that number is not an entry on a leaderboard somebody else runs. It is a value inside a comparison table Perceptron built by running the models in its own environment. In that same table GR00T N1.7 reads 97.0% and π0.5 reads 96.9%. All three sit inside 0.3 percentage points. Before saying which model is better, you have to say that this yardstick can no longer tell the three apart.

ROBOTERA VPP2's "open weights" is also hard to flatten into one line. The official organization account holds a single example checkpoint under non-commercial terms, while the weights actually in use sit in a personal account that the official code repository points to. That one rests on an Apache 2.0 base model. A single sentence saying "we released it" changes what you are allowed to do depending on which of the two it means.

5.3Figure Helix 2.5 Wrote Down Its Attempt Counts Off the Board

One of the top seven has a ○ in the evaluation-conditions column: Figure Helix 2.5 at 6th. The company rented thirty homes in the Bay Area to test in, and wrote down the raw figure of 237 successes out of 420 attempts. It even disclosed that no partial credit was given. It simply did not go up on a shared board. On the measure of recording how many tries were made and how many worked, it does better than most rows on the board. Writing down the conditions and turning up on a shared board are two different acts, and few places do both.

The other six are not silent on conditions either. Gemini Robotics 2 gave success rates by task group as ranges, Skild AI recorded a conversion in which one video is worth roughly 380 demonstrations, and Generalist wrote that the average across ten tasks rose from 59% to 83%. Dyna put mean time between interventions forward as its metric, and Dream-1 stated that it works on twenty-seven robot forms it was never trained on. These numbers do share one thing, though. Every one of them sits on tasks the company built itself, which means they cannot be compared with each other. Why robot training data and interfaces are so scattered is something Pebblous took up separately in why robot training data is both abundant and insufficient.

Open your weights and it also becomes possible to count whether people actually pick them up. Hugging Face records which model a checkpoint was built on, so the number of derived checkpoints resting on a given model can be queried in reverse. That is one of the few objective signals for whether an opened model gets reused. On 2 November 2026 we queried the public repositories of the fifteen one by one.

The Hugging Face API was queried directly on 2 November 2026. Derived checkpoints come back capped at 1,000 per request, so SmolVLA hit the cap and the real figure is higher. Downloads are the 30-day running total. The bottom three rows are a comparison group that is not on the LinkedIn list. ⚠️ The 0 on OpenGalaxea/G05 is handled separately below.

The first thing that catches the eye in the table is the contrast between the top row and the two at the bottom of the on-list group. GR00T N1.7 has 298 derived checkpoints; Perceptron Isaac 0.5 has zero. Yet Isaac 0.5 is the one with the looser terms: Apache 2.0 and no access restriction. GR00T sits behind a company license plus one more terms acceptance for the backbone. The side with the lower threshold goes unused, and the side with an extra door was used 298 times. Opening something, by itself, does not get priced.

What does get priced is not the license but everything around it. GR00T comes with integration into training tooling including LeRobot, hardware bundles, examples and documentation. Isaac 0.5 is a 36B mixture-of-experts build, so the cost of simply getting it running is different. This is not an observation that Isaac 0.5 is a bad model. It is an observation about the ecosystem: the most widely opened model among the fifteen sits at the bottom of the list and almost nobody picks it up.

6.1Two Things These Numbers Cannot Measure

Using download counts means writing down two limits first. One: a top list filtered by tag cannot be used as is. The most downloaded entry under the Hugging Face robotics tag is autotrust/JEV-9B at 912,824, and that is a classification model, not a robot policy (Pebblous wrote a separate report testing that model's cost claim). Second place is a video understanding model and fifth is an autonomous driving model. Carry the tag's top entries over as a "robot model ranking" and the sentence becomes false. The table above was filtered by hand down to manipulation policies only.

Two: a 0 is not "nobody downloaded it". The Hugging Face tally counts requests for a particular file, so a repository laid out differently returns 0. That is the case for OpenGalaxea/G05, which carries 22 likes. And checking via the API shows the repository requires accepting terms. A limit in how the tally works and a real access threshold are overlaid here, so a single 0 cannot tell you which one it is. For a row showing 0, likes are the only thing that speaks. And in this report the download count says exactly one thing: that someone outside could get their hands on it.

6.2Korea Appears Only on a Third Board, as RLDX-1

We went through the affiliation field on all 79 board rows and all 33 RoboTwin rows. There is no team identifiable as a Korean institution. None on the fifteen of the LinkedIn list either. One name could cause confusion: the affiliation on PhysicalRSI at 9th on sim reads "HKU MMLab & KAI", and that KAI is the name of a project at a University of Hong Kong lab, unrelated to Korea.

Widen that fact to the Korean ecosystem as a whole, though, and it becomes false. There is RLWRLD's RLDX-1, an 8.1B model announced in May 2026, with eleven repositories up on the company's Hugging Face organization account with no access restriction. In that you do not have to accept terms to download it, the threshold is lower than for several open-weight models on the list. And it sits on another official leaderboard, RoboCasa365, at 15th of 23 entries with an overall score of 36.0, submitted by RLWRLD itself. GR00T N1.5 (17th) and N1.6 (19th) on the same board are both baselines run by the RoboCasa team, and RLDX-1 scores above both. Derived checkpoints number 42, more than every open-weight model on the list except GR00T N1.7. The license is non-commercial only.

And yet the question this report put to AGIBOT and ROBOTERA works here just as well. RLWRLD's official site records its RoboCasa 365 score as 32.1 and the comparison figure for GR00T N1.6 as 26.9. The current values on the official board are 36.0 and 21.9. Both differ. We could not confirm the cause. The board may have re-evaluated since, so we record these only as two snapshots from different dates. We do not conclude that the figures were inflated. The company also promotes having led on eight tasks, but that is a comparison of individual sub-tasks one against another, while on the overall board measuring many tasks at once it is 15th. "First" and "15th" can be true of the same model at the same time because the denominators differ.

The rest of the domestic picture has to be set down briefly. NAVER LABS' DIVINE, Rainbow Robotics' RB-Y1, HD Hyundai Robotics' upper-body model, ROBOTIS' humanoid platform, and demonstration projects from Tomorrow Robotics and NAU Robotics were all announced during 2026, but we found no case with a confirmed public benchmark score or published weights. The ROBOTIS setup is not an in-house model but a configuration running GR00T N1.7. The structural reason robot experience data cannot simply be bought is something Pebblous has covered separately.

6.3Four Questions That Work on Any Supplier's Scorecard

Carry the six columns above outside robotics and they come down to four questions. Not one of them is specific to robot models. Wherever a scorecard made elsewhere has to be turned into a yes or a no, it is the same four.

• What was the number run on? Is it a number on a yardstick someone else built, or on tasks the supplier built itself? If 97.2% and 97.0% are 0.2 points apart, that yardstick can no longer tell the two apart.
• Who measured it? Did the supplier submit it, or did a third party run it? That this splits even inside one table is what section 4 of this report is about.
• Are the conditions written down? How many attempts, what counted as success, whether partial credit applied, whether extra data was used. A figure recorded as 237 of 420 is more useful than a line in a ranking.
• Can we run it again ourselves? Weights or API calls, is there a path to reproducing it on our own tasks? Without one, the score is reference material rather than grounds for a decision.

None of the four is about running models down. They are about making scorecards readable. What happens when a number nobody measured in the field gets adjudicated from documents alone is something Pebblous also took up in a report examining Anthropic's robot automation outlook.

Why This Matters to Pebblous

This Is Not a Review of the Models

This report does not say which robot model is good. The material needed to answer that question is not public, and pretending to answer it would mean doing the same thing the LinkedIn list did. Instead we stepped one square sideways and counted how far the evidence goes. In effect we put the questions Pebblous asks first whenever customer data arrives to a model scorecard: what the number was run on, who ran it under what conditions, whether we can run it again. The six columns built in section 5 are not a robot-model-only form. They are an adoption review form that stands up against any AI supplier, and the four questions in section 6.3 are the instructions for using it.

The Same Thing Happens in Customer Data

First, an aggregated metric hides its population. The most downloaded entry under the Hugging Face robotics tag was not a robot policy, and a value of 0 did not mean nobody downloaded it. In neither case was the number wrong; in both cases what the number was counting had not been written down. Confirming field definitions and the meaning of missing values in customer data before anything else is exactly the same work.

Second, the same name does not mean the same thing. "GO-3 Preview" against "GO-3 Nano Preview", "LingBot-VLA 2.0" against the unversioned "LingBot-VLA", π0.7 against π0.5 — these kept catching throughout this report. What deserves more attention is who dropped that one word: not an outside observer, but the company that made the model. The same-name-different-entity problem DataClinic keeps finding in customer data arises in model names in exactly the same way.

Third, failure only becomes visible when baselines are in the table. Section 4 of this report was possible because RoboDojo ran thirty policies itself at launch and put them in the same table. Without those rows the board would have been a well-dressed list of submissions and nothing else. But RoboDojo left that distinction as a blank while RoboTwin wrote the name in. Recording the same information or leaving it empty changes the conclusion a reader draws. This is the cheapest improvement available to anyone designing a scorecard.

What This Report Counted Was a Gap

Earlier pieces here on benchmarks mostly pulled the instrument apart to show where it was broken. Rankings shifting when benchmarks turned out to overlap, a model built to judge failure doing worse than a general-purpose one, counting the floor of defects in a physics problem. This time what we counted was not the broken place but the empty one: the place where a usable yardstick stands open and the top seven do not turn up on it, and, among those who did turn up, the place where who submitted the row goes unrecorded. A flaw can be fixed. A gap gets filled only when somebody shows up.

The Korean material at the end of section 6 is there for the same reason. Put the question to Chinese and American first-place claims and not to domestic models, and the question stays a tool for aiming at other people. We put it, the same mismatch came back, and we wrote that down as found. The use of this report is not in settling who is first. It is in deciding, before the next ranking table lands, what to ask first.

References

Primary sources — the measuring side

• 1.RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies, arXiv:2607.04434 (v1 2026-07-05, v3 revision 2026-07-08), 44 authors. Five capability axes, 42 sim tasks plus 18 real-robot tasks, 30 policies integrated, RealEval cloud real-robot runs. ⚠️ This report checked only the abstract and metadata. The submission and evaluation protocol section was not opened.
• 2.RoboDojo leaderboard, robodojo-benchmark.com (checked 2026-11-02, snapshot 2026-10-30). The JavaScript bundle that draws the page carries 66 sim rows plus 13 real-robot rows together with 28 changelog entries. The governance quotation, the "(agent track)" tags and the 2026-07-06 list of 30 launch models all come from here.
• 3.RoboTwin 2.0 leaderboard, robotwin_leaderboard.json (parsed 2026-11-02, board.updated = 2026-10-30). 33 rows (cotrain 28 + sft 5). Robot, training data, number of evaluation runs and default submitter are all recorded together in the data file.
• 4.RoboCasa365 official leaderboard (updated 2026-11-01, 23 entries). The RLDX-1 placing of 15th at 36.0 overall in section 6.2 and the GR00T N1.5 and N1.6 rows come from here.
• 5.World-Action Models for Robot Learning and Control: A Survey, arXiv:2609.16074 (2026-09-13, 16 authors). The boundary between VLA and WAM, plus a catalog of the datasets, benchmarks, metrics and protocols used in evaluation. ⚠️ Checked only as far as the abstract.
• 6.What Matters in Designing World Action Models: An Empirical Study, arXiv:2609.24048 (2026-09-21). A controlled comparison of 6 causal structures × 8 latent representations × 4 training objectives across three benchmarks (RoboCasa-GR1, LIBERO, LIBERO-Plus), with the key findings validated on DROID real-robot data. The comparison was not run across four.

The LinkedIn post and first-party announcements

• 7.LinkedIn post "Which robot action models are best at learning new tasks?" (lnkd.in/p/griDTpn4), 2026-10-31 16:31 UTC. The posting time was computed from the activity identifier. ⚠️ This report deals with the content of the post and not with its author. Whether the post was later edited or deleted is unconfirmed.
• 8.embodied.guide (checked 2026-11-02). A catalog of 366 pieces of robot hardware. The quotation "Independent by design — No paid listings. No affiliate links. No manufacturer rankings." is from here.
• 9.ROBOTERA press release (2026-10-09, PR Newswire, datelined Beijing) — "Tops Embodied AI Benchmark RoboDojo Without Additional Data or Agent RSI". VPP2 at 32.26% average success rate, 39.26 average score, 45.0% on LIBERO-Pro. The video prediction model weights were released the following day.
• 10.AGIBOT Innovation Day 2026 (2026-10-29, Shanghai). Announced that GO-3 Preview held first place on RoboDojo, RoboTwin and RoboColiseum. ⚠️ The board row is named "GO-3 Nano Preview", and no public leaderboard for RoboColiseum could be found.
• 11.Bundle of first-party announcements for the fifteen — Skild AI S1 (2026-08-25) · Generalist GEN-1.5 (2026-08-19) · Gemini Robotics 2 (2026-07-30) · π0.7 (2026-04-16) · Dream-1 (2026-10-19) · Figure Helix 2.5 (2026-09-17) · Dyna-2 (2026-08-10) and Dyna-2.1 (2026-09-29) · AGIBOT GO-3 Preview (2026-10-29) · X Square Robot WALL family · LingBot-VLA 2.0 · GR00T N1.7 (2026-04-17) · ROBOTERA VPP2 (2026-10-09) · Unitree UnifoLM-WLA-1.0 (2026-09-10) · Galaxea G0.5 (2026-08-12) · Perceptron Isaac 0.5 (2026-08-26). The monthly count in section 1.1 and every cell of the table in section 5 came out of this bundle.
• 12.Related papers — Galaxea G0.5 (arXiv:2608.11739) · LingBot-VLA 2.0 (arXiv:2607.06403) · the GR00T lineage (arXiv:2503.14734) · Gemini Robotics 1.5 (arXiv:2510.03342).

Model cards and direct measurements

• 13.nvidia/GR00T-N1.7-3B — the sentence "This model is released under the NVIDIA Open Model License Agreement." appears verbatim in the README. The backbone nvidia/Cosmos-Reason2-2B is a repository requiring acceptance of terms.
• 14.Derived checkpoints and downloads measured directly (Hugging Face API, 2026-11-02) — PerceptronAI/Isaac-0.5 · robbyant/lingbot-vla-v2-6b · OpenGalaxea/G05 · unitreerobotics/UnifoLM-WLA-1.0-Base · x-square-robot/wall-oss-0.5 · lerobot/smolvla_base · lerobot/pi05_base · openvla/openvla-7b. Derived counts were taken by reverse lookup on base_model:, and the response headers were checked for truncation.
• 15.arXiv search API monthly counts (2026-11-02) — within cat:cs.RO, papers whose abstract contains the phrase "vision-language-action" rose from 74 in 2026-04 to 276 in 2026-09, and "world action model" from 8 to 90. ⚠️ A crude phrase-match count.
• 16.RLWRLD RLDX-1 — 11 repositories on the Hugging Face organization RLWRLD (no access restriction), 42 derived checkpoints on RLDX-1-PT, RLWRLD Model License v1.0 (non-commercial). The discrepancy between the RoboCasa 365 figures on the company's own page and the current values on the official board was established by opening both sources on the same day.

Neighboring Pebblous reports

Three Teams, Three Robot Brains (2026-04-24) — the same ground seven months ago. · Collapsing Two Benchmark Pairs Moved 22 of 51 Physical AI Models (2026-08-28) — the story of benchmarks overlapping. · Models Built to Catch Robot Failures Lost to Off-the-Shelf VLMs (2026-09-05) · Anthropic Grades Robots Without Watching a Single One (2026-10-03) · Fusion Reactor Dataset BoSS Grades AI Designs by Rerunning the Physics (2026-11-02) · Why Is Robot Learning Data Still Scarce When Video Is Everywhere? (2026-08-11) · Can an AI That Only Picks Answers Really Work This Cheap? (2026-09-27) · Robot Experience Data Can't Be Bought (2026-07-20) · Would that score hold up on your own codebase? (2026-09-15).