AXIS Dataset V1: continual pretraining lift on LIBERO-Plus. Infrastructure is only as credible as the signal it puts into models. AXIS Dataset V1 is the public empirical check on that signal: continual pretraining on V1, evaluated on LIBERO-Plus, with π0.5 as the reported policy family. Result (as stated in the Axis brief): Continual pretraining on AXIS Dataset V1 lifts π0.5 from 83.9% to 88.8% on LIBERO-Plus. That is a measured change under a fixed evaluation, not a claim that every crowdsourced corpus will move every benchmark. The useful reading is narrower: data produced through Axis’s collection and processing path can improve a strong baseline when used as continual pretraining fuel. How to read the number: - What moved: pretraining exposure to V1, then LIBERO-Plus score for π0.5. - What the score is: task success under the LIBERO-Plus protocol used in the report (not a general “robot IQ”). - What it does not prove alone: sim-to-real for every embodiment, or that scale without structure is enough. Where the brief notes scaling behavior, performance rises as a larger fraction of V1 enters pretraining, without an obvious early plateau in the published curve. That pattern matters for systems design: the loop is built to grow accepted, structured trajectories, then test whether more of that corpus still buys lift. V1 sits downstream of the product surfaces in Note 2 and can populate Task Packages (Note 3). Upstream proof that the loop can close on hardware was Little Prince (Note 1). This note only records the benchmark-side evidence. Sim Dataset V2 and broader pretrain regimes are later extensions on the same measurement habit: ship data, report lift, bound the claim. @axisrobotics

Task Package as the commercial unit of robot data infrastructure A pile of trajectories is not a contract. Partners need a unit that states what skill coverage was bought, under what variation, for which scenario class. On Axis, that unit is the Task Package. It is not “a zip of demos.” It is the core commercial unit: a structured delivery whose value is measured across four dimensions. 1. Scenario: The environment and goal class: objects, layout families, success criteria, and the operational context the policy must handle. 2. Atomic skills: The primitive actions the package is meant to support (reach, grasp, place, pour, open, and so on), composed into the task rather than left as an undifferentiated blob of motion. 3. In-task randomization: Controlled variation inside the scenario: poses, object instances, lighting, distractors, ordering. This is what turns a single demo path into coverage useful for training and transfer. 4. Trajectories: Volume and acceptance quality of the recorded runs that populate the package. Count matters only after checks and processing, raw session volume is not the SKU. Delivery form can change (static dataset, continual prior, few-shot or corrective loop). The unit of account does not: partners still buy or specify a Task Package along those four axes. That design ties back to the loop in the previous note. Task generation instantiates packages, collection and processing fill them, deployment and failure signal which dimension to expand next (more skills, harder randomization, new scenarios). Task Package is how infrastructure becomes something you can scope, price, and repeat. Benchmark lifts and operating scale are separate notes. @axisrobotics

Doma Weekly: Frontier's First Vault, Product Updates and Growth Keeps Climbing. 💠 First vault on Solana: Frontier’s first vault went live on Solana through Loopscale, an independent third-party vault provider. Capital sits with Loopscale. Frontier only records points and extension priority. The flow has three steps: 1. Deposit eligible assets into Loopscale. 2. Link the Loopscale wallet to a Frontier account. The longer the deposit sits, the more Frontier Points it accrues. 3. Commit those points to the extensions you want to back: .agent, .crypto, .sol, .wallet, .robot, .human, and others. Frontier figures at publication: TVL 183K, members 43K+, extensions live 148. 💠 Product: The ship was about data consistency and unread signals, not a new trading tool. - The Explore tokens table gained a sortable launch-date column, part of consolidating the table onto a single domains data source. - The notifications bell now has a red unread-count badge. The DM icon has a red dot for unread messages. - Table headers and rows no longer scroll out of sync at the horizontal edges. - Large percentage-change values no longer overflow the homepage “Just Launched” cards. - The domain bonding panel moved onto the shared progress-bar helper. - Additional bugfixes and stability work across the app. 💠 Protocol numbers: All figures are cumulative: - Volume: $486.01M. - Transactions: 41,911,158. - Tokenized assets: 258,561. - Wallets: 59,554. - Domains launched: 756+. 💠 Community and calendar: The Solquicks × Doma AMA on 30 September, on Frontier’s meaning for Solana, drove the week’s chat spike of about 2.5K messages, the biggest week yet by message count. The midweek walkthrough covers four claimable earning types: launch proceeds, launch pool fees, position fees, and vested tokens, and the split between locked, circulating, and unlocked. ICANN set Reveal Day for 7 October 2026, when it publishes proceeding gTLD applications: primary strings, variants or replacements, contention sets, and the public portion of each filing. For Frontier, that is the first official disclosure date the extension-backing layer can be checked against. Read details: blog.doma.xyz/doma-weekly-fr… @D3inc @domaprotocol

Four product surfaces, one compounding data loop Axis is best described as full-lifecycle robot data infrastructure, not a single dataset product. Four surfaces sit on one loop: each stage conditions the next. 1. Task generation Turns a data requirement into scalable task families: scenarios, objects, layouts, success criteria, and acceptance checks. This is the specification layer. Without it, collection has no contract. 2. Simulation / browser collection Contributors teleoperate simulated robots in the browser. No local simulator. No on-site hardware. The interface supports large-scale human demonstrations for pretraining and, where designed, human-gated correction during policy rollouts for post-training. 3. Ego-centric / mobile capture Complements sim with first-person, in-the-wild human activity data. Same goal: structured trajectories, different embodiment and diversity profile. 4. Data-to-model pipeline Filter, replay, smooth, and package trajectories into training-ready form; train; deploy; route failure and correction signal back into what should be collected next. The loop Specify → collect → process → train → deploy → correct → specify again. What compounds is coverage (tasks, skills, randomization) and the match between failure modes and new data. What does not compound is undifferentiated volume: sessions only matter when they pass acceptance and enter a defined training path. The four surfaces are not four disconnected tools. They are interfaces on one system: task families in, policies and feedback out. Commercial delivery later sits on the same structure (Task Package as the unit of account). Architecture first; packaging second. This note only maps the system. Measured lifts (Dataset V1) and operating scale are separate notes. @axisrobotics

Little Prince: closing the loop from browser teleoperation to real-robot policy Little Prince (Axis’s rose-watering experiment) was a controlled end-to-end run of the data path: task definition in simulation, large-scale browser teleoperation, processing, training, and deployment on a physical robot. Contributors teleoperated a simulated robot in the browser. No local simulator. No on-site hardware. Acceptance depended on task checks; only valid sessions entered the training set. Measured scale: 15,371 participants produced 85,387 sessions. From collection through training to real-robot deployment took five days. The sequence: 1.Specify the task (objects, layout, success criteria). 2.Collect trajectories through a web interface under acceptance rules. 3. Process (filter, replay, smooth) into training-ready data. 4. Train a policy on that corpus. 5. Deploy on hardware and observe execution. In older setups, demonstration data is often gathered in small labs over weeks. Here, collection was parallelized through the browser, and the full generate → collect → train → deploy chain was exercised on one task family within a short window. This does not claim that every crowdsourced run will transfer. It shows that, under this design, community teleoperation can supply structured trajectories sufficient to reach a real-robot policy on a defined timeline. Little Prince is best read as an existence proof for the loop: participation matters only insofar as it becomes accepted trajectories, and those trajectories feed the same infrastructure that produces the next task delivery. @axisrobotics

From One Prompt to Infinite Tasks Physical AI needs more than a large number of trajectories. It needs training tasks with enough variation to cover different physical situations. This is the problem Axis approaches with its Task Generation Engine. The concept is: Prompt → Infinite Tasks Give the engine a scene, a set of objects, and a goal. It can automatically: Select assets. Compose scenes. Sample layouts. Configure physical conditions. These inputs can then produce task families that vary across: Scenarios. Object types. Spatial layouts. Embodiments. Visual conditions. Axis defines the commercial unit built from this process as a Task Package: Scenario × Skills × Randomization × Trajectories. Each dimension represents a different part of the data package. Scenario defines the physical environment and deployment context. Skills define the manipulation primitives, such as: Grasp. Place. Push. Rotate. Randomization provides variation through spatial layouts and asset configurations. Trajectories represent the volume of usable training data ultimately delivered. The brief gives a simple example of in-task randomization: 40 layouts × 4 asset variants = 160 instances The important point is not the number 160 by itself. It is the mechanism behind it: Layout variation + Asset variation → More task instances This is how Axis approaches diversity at the task-generation layer. Instead of treating data collection as repeatedly performing the same fixed task, the system first creates structured variation in the tasks themselves. Then those tasks can feed into Axis' distributed browser teleoperation network to generate trajectories. So scaling Physical AI data has two sides: Generate more tasks. Collect more trajectories. Axis is building the infrastructure connecting both. @axisrobotics

Doma Weekly not a small-ship week. Three signals at once: product, protocol, allocation. 1/ Product New start flow: app.doma.xyz/start Two doors: trader and domain owner. Account panel now leads with Swap; Receive merged into Deposit. Friction cuts: chart double-init, claim URL after refresh, missing fractional records blocking bridges. This week reduced the cost of understanding the product. 2/ Protocol - Volume: $458.93M (closing in on $460M). - Transactions: 40,599,281. - Tokenized assets: 256,003. - Wallets: 58,916. - Domains launched: 747. The secondary book is compounding. Track this before the slogans. 3/ Frontier: new primitive. ICANN 2026 TLD round is in flight (1,600+ applications). Last comparable round: 2012. Frontier is priority access to new extensions (.agent, .robot, .human, .wallet, .nft, .sol, and others). Mechanism: Deposit into a supported vault, earn yield plus Frontier points, stack points on the TLDs you want. At launch, committed points decide who picks a name first. Signup: +10,000 pts. Epoch 1: 4x, through 21 Mar 2027. Deposits open 28 Sep. Members were 25K+ at weekly publish, product is already ~29K. This is not points-for-points. It is a ranking system for scarce namespace. Why the week matters: Doma is now operating two markets: fractional trading of existing domains, and pre-launch allocation for domains that have not been issued yet. Onboarding lowers friction on market one. Frontier builds order flow for market two. Read details: blog.doma.xyz/doma-weekly-fr… @domaprotocol @D3inc

Why Video Isn't Enough for Physical AI The internet has an enormous amount of video. But video alone is not enough to teach robots how to act in the physical world. The problem is ground physics. Axis identifies a specific limitation of internet video: It does not contain joint torques, contact forces, or 6-DoF poses. A model trained only on pixels can learn what an action looks like. But Physical AI needs data that captures the physical interaction behind that action. This is one of the reasons Axis chose simulation as a core method for data collection. Instead of relying on internet video, Axis provides a Simulation Data Collection Platform where contributors can teleoperate simulated robots directly through a browser. No local simulator. No specialized hardware. The platform supports two workflows: Human demonstrations → large-scale pretraining Human-gated corrections → post-training with DAgger This creates a way to collect structured robot interaction data at scale while keeping the collection process accessible through the browser. The broader Axis data pipeline then takes those raw trajectories and processes them into training-ready data through validation, filtering, smoothing, resampling, and replay under randomized conditions such as cameras, lighting, textures, object poses, mass, and friction. So the problem is not simply: “We need more robot videos.” It is: “We need physical action data with the information required for robot learning.” That distinction explains why Axis is building around simulation, teleoperation, and a full data processing pipeline. For Physical AI, seeing the world is only part of the problem. The harder problem is collecting data that teaches machines how to act within it. @axisrobotics

Doma Weekly recap: The interesting part of this week is not any single ship. It’s that the buy path, the account surface, and the agent loop are starting to sit in one system. Product Updates: - Domain buyout now has its own UI on the token page. Purchase is no longer jammed into the same view as everything else. - Owned domains live in the account side panel. You can scan inventory without leaving the page. -Community threads expand inline. Context stays on the token. - Tag doma-ai in chat for price or token info. - Gochujang pack opens now credit Doma Points automatically. - Terminology is finally split: Token FDV, Token price, Domain sale price. - Swap fee breakdown is back. LP withdrawal warning is clearer. Swap/strategy bugs got cleaned up. Why that matters: DomainFi only works if “token” and “domain” stop being used as synonyms. Liquidity, valuation, and buyout are different objects. The UI is starting to treat them that way. By the numbers: - Volume: $424.61M. - Transactions: 38,271,523. - Tokenized assets: 253,939. - Wallets: 57,830. - Domains launched: 743+. Gochujang.com Still the loudest cultural loop on the network. - 1.4M packs opened. - 950+ holders. - $34M+ volume. - $855K token FDV. Packs still mint real GOCHUJANG domain tokens plus points. That is the point: a game that settles into inventory, not just a leaderboard. Agentic layer Inder Singh (CTO, D3) on Stabledash: “It’s not humans versus AI, it’s humans driving AI.” The conversation went from revenue management into where agents actually attach next, domains, checkout, personal assistants. If agents start booking, bidding, and managing names, the domain is no longer a static asset. It becomes an address for machine commerce. Next Update: - New DOMAXX cohort Monday. - KBW 2026 and the ICANN TLD cycle are the two calendars that matter from here. New gTLDs do not come around often. The last landrush was 2012. Full Details Blog: blog.doma.xyz/doma-weekly-pr… @D3inc @domaprotocol

Another week of steady expansion for Doma. The interesting part is not just the numbers. It is how the product layer around DomainFi keeps getting deeper. This week: - Grid Trading is now live, adding another automated strategy designed for choppy, range-bound markets. - Doma reached $379.77M in total volume, 35.6M+ transactions, 252K+ tokenized assets, 57.7K+ wallets and 725+ domains launched. - Gochujang.com keeps gaining traction with 558K packs opened and 1,132 token holders. - Pack pulls now earn Doma Points too, turning activity inside the creator-coin game into another layer of ecosystem participation. • The latest D3 discussion also went deeper into ICANN's next TLD round, fractional domain ownership and what the next generation of domain infrastructure could look like. The broader pattern is becoming clearer: Doma is building more than a marketplace for tokenized domains. Trading, ownership, creator economies and domain infrastructure are gradually converging into one DomainFi stack. And the product is getting deeper alongside the growth. @domaprotocol @D3inc