← the graph
Sector 2 · 6 sub-themes

World Models & Simulation

Dyna Robotics' Dyna-2 is the most consequential release here: a World Action Model pre-trained on over one million hours of first-person human video — zero robot data — that still outperforms models trained on actual robot demonstrations. The core insight is that human egocentric video enables cross-embodiment generalization better than robot-specific teleoperation data, reframing the entire robotics data bottleneck away from hardware and toward large-scale human experience capture. Researcher Jason Ma frames this as robotics' first true scaling law: more human video produces predictable, consistent performance gains. That finding puts pressure on the sim-data approach simultaneously being scaled elsewhere — Google DeepMind's Gemini Robotics 2 uses sim-only post-training for contact-rich tasks, while Axis Robotics has crowdsourced 3 million trajectories via 123,000 contributors, tripling its dataset in weeks. DeepSeek's capital partnership with Unitree Robotics signals that frontier LLM labs are now moving directly into physical AI, betting that language-model-scale reasoning transfers to humanoid bodies. On the generative side, LTX-2.5 brings open-weights world modeling to local deployment, generating 10 seconds of video from a single image in under 7 seconds — relevant both for creative applications and as simulation infrastructure. China's Fengwu, Pangu, and Fuxi models, deployed in real-time during Typhoon Dolphin tracking, show world models earning operational trust in high-stakes forecasting.

Dyna RoboticsDyna-2Jason MaGoogle DeepMindGemini Robotics 2Axis RoboticsDeepSeekUnitree RoboticsLTX-2.5FengwuPanguFuxiWorld Action Modelegocentric human video pretrainingcross-embodiment generalizationrobotics scaling lawsphysical AIcrowdsourced robot trajectoriessimulation-based trainingopen-weights world modelingAI weather forecastingfrontier LLM to humanoid transfer
2.1

Simulation Training and Crowdsourced Data Scale Robotic AI Development

  • Google DeepMind's Gemini Robotics 2 uses sim-only post-training on a mixed-data base model to handle automotive kitting and other contact-rich tasks, meaning the robot's post-training phase relies entirely on simulated environments rather than real-world data collection. [1]
  • Axis Robotics' robotic data engine on the Base platform scaled from 1 million trajectories on June 7 to 3 million trajectories, a 3x increase, generated by over 123,000 contributors worldwide. [2][3]
  • A widely noted challenge in physical AI development is that robots require large volumes of high-quality, diverse data to improve, and static datasets are insufficient because real-world variation demands continuous, dynamic data collection. [4]
Google DeepMindGemini Robotics 2Axis RoboticsBaserobotics AI trainingsimulation-based post-trainingrobotic data enginessynthetic trajectory generationcontact-rich manipulation tasksphysical AI developmentcrowdsourced robot datareal-world data scarcityautomotive kittingdynamic data collection
Sources
  1. [1]axisrobotics — Gemini Robotics 2 by @GoogleDeepMind uses sim-only post-training on a mixed-data base model to handle automotive kitting. Contact-rich tasks★twitter
  2. [2]mdshefat217 — 3M robotic trajectories on @base. @axisrobotics went from 1M on June 7 to 3M today. What caught my attention isn't just the number. It’s howtwitter
  3. [3]axisrobotics — Our robotic data engine on @base just hit 3M trajectories. 1M on June 7. Now 3x. Generated by 123k+ contributors worldwide, with contributiotwitter
  4. [4]alveejack1 — Physical AI has one big challenge people often miss: robots need a lot of good data to improve. Static datasets are not enough because real twitter
2.2

DeepSeek forms capital partnership with Unitree Robotics for physical AI

  • DeepSeek, the Chinese AI startup, is entering a capital and business partnership with Unitree Robotics (宇樹科技), a humanoid robot developer, with the two companies cooperating toward the realization of AGI with human-equivalent capabilities. The deal was revealed through IPO-related documents disclosed by Unitree, which show DeepSeek investing 140 million (figure truncated in the paywalled source) in conjunction with Unitree's IPO. [1][2][3][4][5][6]
DeepSeekUnitree RoboticsAI-robotics partnershiphumanoid robotsAGI developmentChinese AI startupcapital investmentIPO fundinghuman-equivalent capabilities
Sources
  1. [1]nikkei — DeepSeek、ヒト型ロボ新興と資本提携 中国でフィジカルAIを推進 https://t.co/CnZxGtnigp★twitter
  2. [2]4ki4 — DeepSeek、ヒト型ロボ新興ユニツリーと資本提携 中国でフィジカルAIを推進 - 日本経済新聞 https://t.co/9PbBr7qEtS★twitter
  3. [3]SOU_BTC — DeepSeek、ヒト型ロボ新興と資本提携 中国でフィジカルAIを推進 - 日本経済新聞 中国のDeepSeekはヒト型ロボット開発企業、宇樹科技と資本業務提携する。AIとロボット分野で中国を代表する両社が、人間に匹敵する能力を持つ汎用人工知能（AGI）の実現に向け協力する。 htwitter
  4. [4]nikkei_business — DeepSeek、ヒト型ロボ新興と資本提携 中国でフィジカルAIを推進 https://t.co/GoMk5XF1GCtwitter
  5. [5]nikkei — DeepSeek、ヒト型ロボ新興と資本提携 中国でフィジカルAIを推進 https://t.co/CnZxGtnigptwitter
  6. [6]nikkei_business — DeepSeek、ヒト型ロボ新興と資本提携 中国でフィジカルAIを推進 https://t.co/GoMk5XF1GCtwitter
2.3

Dyna Robotics releases Dyna-2 World Action Model trained on 1M hours

  • Dyna Robotics released Dyna-2, a World Action Model (WAM) explicitly distinguished from Vision-Language-Action (VLA) models; it uses video diffusion as its base architecture to jointly learn future video prediction and actions rather than treating them separately. [2][3][4]
  • Dyna-2 was pre-trained on over one million hours (described as approximately 170 years) of human first-person activity video, covering tasks such as cooking, tidying, and laundry folding. [5][6][4]
  • Alongside the model release, Dyna Robotics introduced a laundry-folding robotics benchmark, framing it as a standardized evaluation task for the broader robotics research community. [1]
Dyna RoboticsDyna-2World Action ModelVision-Language-Actionvideo diffusionrobotics foundation modelworld action modelvideo diffusion architecturefuture video predictionrobot action learningfirst-person activity video pretraininglaundry folding benchmarkembodied AI evaluation
Sources
  1. [1]shreyasgite — Dyna made laundry folding into a standard Robotics Benchmark, now lets see what this leads to! So a few things to ponder from first read: - twitter
  2. [2]stevencheng — Dyna-2, a World Action Model (WAM) — not a Vision-Language-Action (VLA) model — pre-trained on one million hours of human activity data. WAMtwitter
  3. [3]LeoKharon — NEW MODEL: Dyna scales up to 1M hours worth of training! You probably have seen it, but it's worth dicussing: Dyna Robotics' new human-videotwitter
  4. [4]OsoneHiroyuki — Dyna Roboticsが「Dyna-2」を公開。 Dyna-2はVideo Diffusionをベースに、未来動画とActionを学習するWorld-Action Model（WAM）。100万時間超の人間の一人称作業動画で事前学習している。 データは料理、片付け、衣類折り、twitter
  5. [5]openkachi — so @dynarobotics just put out a robotics paper that's kind of hard to ignore. the short version: they trained a model on over a million hourtwitter
  6. [6]Anstronaut1 — DYNA-2 World Action Model launches from Dyna Robotics 🤖 ▪️ Trained on 1M+ hours human first-person video (170 years of experience) ▪️ Max 90twitter
2.4

Dyna Robotics' Dyna-2 trained on 1M hours human video, outperforms robot-data models

  • Dyna Robotics released Dyna-2, a World-Action Model pre-trained on 1 million+ hours of egocentric (first-person) human video with zero robot data used during pre-training, and claims it outperforms models trained on actual robot data. [2][5][6][9]
  • Researcher Jason Ma describes Dyna-2 as establishing the first true scaling law in robotics: robots trained on increasing amounts of first-person human video show predictable, consistent performance improvements. [4][10]
  • The central finding is that human video data's primary benefit is enabling cross-embodiment generalization — the ability to transfer learned physical understanding across different robot bodies and novel environments without robot-specific training data. [1][7][11]
  • The Dyna-2 results suggest the robotics data bottleneck is not hardware or robot teleoperation volume, but access to large-scale human experience data, reframing where the field should focus data collection efforts. [3][5][8]
Dyna RoboticsDyna-2Jason MaWorld-Action Modelegocentric human video pretrainingrobotics scaling lawscross-embodiment generalizationrobot data bottleneckfirst-person video for roboticsembodied AI training datazero robot data pretraining
Sources
  1. [1]Jesse_Y_Zhang — “[Human] video data’s benefit…is cross-embodiment generalization.” Dyna-2’s great blog post finds that scaling action-free human video alonetwitter
  2. [2]0xrichboy — What if the fastest way to teach robots isn't more robots, but more human experience? Dyna Robotics just introduced DYNA-2, a World-Action Mtwitter
  3. [3]DataChaz — This is a fascinating pivot. If Dyna-2 can achieve better robot performance just by watching millions of hours of human video, why are we sttwitter
  4. [4]MTSlive — Jason Ma reveals the first true scaling law in robotics: robots trained on 1 million hours of first-person human video predictably get bettetwitter
  5. [5]LxKus — Hardware is just the vessel. Data is the actual bottleneck. @DynaRobotics just proved it with the Dyna-2 drop. Pre-training on 1M hours of htwitter
  6. [6]VaderResearch — Can robot intelligence scale with video? Dyna-2 trained on 1M hours of egocentric video. ✸ No robot data in pre-training. ✸ Models scale dattwitter
  7. [7]Rana_kamran43 — The interesting thing about Dyna-2 isn’t just the model. It’s the data. Dyna-2 was trained on 1M+ hours of human video and can transfer thattwitter
  8. [8]0xfrigg — the important implication here is not that robot data suddenly matters less it is that physical AI data is starting to split into different twitter
  9. [9]heyrobinai — wait this can't be real a robot model trained on purely human video is outperforming models trained on actual robot data.. with zero robot ttwitter
  10. [10]Anabiya_live — Dyna-2 just put a strong data point on the board: large-scale human video + the right world-action objective can produce predictable gains ttwitter
  11. [11]darshal_ — What caught my attention with Dyna-2 isn’t the 87% number. It’s what’s underneath it. Getting zero-shot performance across novel environmenttwitter
2.5

LTX Releases LTX-2.5 Open-Weights AI World Model for Local Deployment

  • LTX released LTX-2.5, an open-weights AI world model that can be run locally, deployed on-premises, fine-tuned, and integrated into custom workflows, targeting creative, real-time, and physical AI applications. [1][2][3]
  • LTX-2.5 upgrades prompt understanding, motion quality, and multishot consistency compared to its predecessor, and is described as predicting "the next moment, not the next word." [2][4][6]
  • According to source [4], LTX-2.5 can generate 10 seconds of video from a single image in 6.8 seconds, and introduces a new compression method alongside other architectural improvements. [4][5]
LTXLTX-2.5open-weights video generation modelAI world modelreal-time video synthesisimage-to-video generationon-premises deploymentmotion quality improvementmultishot consistencyvideo compressionphysical AI applicationsfine-tunable generative model
Sources
  1. [1]AIFrontliner — LTX just released LTX-2.5, an open AI World Model you can run locally, deploy on-prem, fine-tune, and integrate into your own workflow. Its ★twitter
  2. [2]hasantoxr — LTX just released LTX-2.5, the next generation of its open-source AI world model. It upgrades prompt understanding, motion, multishot consistwitter
  3. [3]NVIDIARTXSpark — LTX-2.5 is out! An efficient, controllable, open-weights world model for creative/realtime/physical AI with higher-quality output, multishottwitter
  4. [4]dr_cintas — Open-source AI just got a real World Model 🤯 LTX-2.5 predicts the next moment, not the next word - 10s from one image in 6.8s: - New comprestwitter
  5. [5]dr_cintas — Open-source AI just got a real World Model 🤯 LTX-2.5 predicts the next moment, not the next word - 10s from one image in 6.8s: - New comprestwitter
  6. [6]hasantoxr — LTX just released LTX-2.5, the next generation of its open-source AI world model. It upgrades prompt understanding, motion, multishot consistwitter
2.6

China's AI weather models challenge traditional typhoon forecasting systems

  • Chinese AI weather models Fengwu, Pangu (Pangu-Weather), and Fuxi represent a new generation of forecasting systems that are challenging the established global forecasting order, working alongside traditional meteorological methods to predict typhoon landfall times and locations with high accuracy. [2][3][4]
  • Specifically in the context of Typhoon Dolphin tracking, the new generation of Chinese AI weather models was deployed in real-time alongside traditional forecasting systems to project the storm's path toward China. [3]
  • China's AI weather forecasting initiative is described as part of a broader response to intensifying extreme weather events, encompassing not only typhoon tracking but also AI-powered flood forecasting and smart reservoir management, supplemented by phased-array radar networks and ground-based remote sensing systems. [1][5]
FengwuPangu-WeatherFuxiTyphoon DolphinAI weather forecastingtyphoon trackingextreme weather predictionChinese meteorological AIflood forecastingsmart reservoir managementphased-array radartraditional vs AI forecasting systems
Sources
  1. [1]globaltimesnews — From #AI-powered flood forecasting to smart reservoir management, phased-array #radar networks and ground-based remote sensing systems, a netwitter
  2. [2]ReutersAsia — China is using AI-powered weather models to track typhoons, with recent forecasts predicting landfall times and locations with high accuracytwitter
  3. [3]Reuters — As meteorologists tracked Typhoon Dolphin's path toward China in recent days, a new generation of AI weather models worked alongside tradititwitter
  4. [4]Calcalistech — China’s AI weather models are challenging the old forecasting order. Fengwu, Pangu and Fuxi are among a new generation of AI systems that catwitter
  5. [5]osint42 — 🔴 China is investing in AI weather forecasting technology as extreme weather events intensify: • The initiative aims to improve weather predtwitter