The Briefing · Archive
All humanoid foundation-model news curated by the humanoid.guide desk — world models, VLAs, training data and compute, with sources linked on every story.
← Back to the reportShenzhen's DexForce trains humanoids in simulated workplaces with varied arm angles, table colours and objects, then transfers the skills to real machines, which CEO Jia Kui told Reuters beats collecting human demonstrations by hand. The company, which Reuters says raised a 1 billion yuan Series B at a valuation above 10 billion yuan, argues the approach could speed a broader robot-capability breakthrough.
NEURA Robotics has introduced Neuron AI, a multimodal foundation model that, according to NEURA, combines 3D vision, touch, audio, robot state, language and memory in a Perception-Language-Action architecture shared across five NEURA robot types. NEURA says it will demo the 4NE1 humanoid running an eleven-minute laundry sequence at NVIDIA GTC Berlin; it has not published benchmark results.
Amazon has launched Physical AI Toolchain on AWS, an open-source stack covering synthetic data, model training, simulation, edge deployment and continuous improvement, integrating NVIDIA Isaac Sim, Isaac GR00T and Cosmos. According to AWS, customers were spending too much engineering effort on infrastructure rather than innovation.
A month after opening its robot training-data marketplace, Kinetic Blocks says roughly nine in ten sign-ups were sellers, which it reads as data production outpacing how robot-policy teams buy data. Its answer is an informal "interests" feature where buyers can flag needs such as egocentric home footage, UMI gripper demos or stereo manipulation data before making a formal request.
Innodata has opened a motion-capture lab in Ridgefield Park, New Jersey, using Vicon optical cameras to record human, robot and object movement directly in 3D rather than inferring it from 2D video. The company will sell off-the-shelf and custom motion datasets, retarget motion across robot platforms, and offer independent robot evaluation and safety validation.
DYNA Robotics has launched DYNA 2.1, a wheeled, two-armed semi-humanoid driven by its DYNA 2 world action model, which the company says was trained on one million hours of human and robot data and paired with a vision-language orchestrator and whole-body controller. According to DYNA, the system completes full commercial workflows such as a laundry shift, and the company judges it on mean time between interventions rather than single-task success; it says robots are already deployed in hotels, laundromats and restaurants.
Image © Humanoid Guide — China Mobile has released Open-RAIL under an Apache 2.0 license, a deployment layer that runs observation, model inference and motor control as separate asynchronous loops to cut the jerky motion of humanoids driven by vision-language-action models. It supports models including Nvidia's GR00T, Physical Intelligence's π0 and AgiBot's GO-1 on robots such as the AgiBot G1, and the team reports large gains in smoothness and task success, figures humanoid.guide notes are not independently verified.
Image © Light Origins — Singapore's Light Origins, founded by former OpenAI researcher Roger Jiang, has launched Light-O1, a 4-billion-parameter embodied foundation model pretrained on up to roughly 100,000 hours of 3D human actions recovered from internet video. According to the company, prediction error on its in-house LightBot humanoid and a Unitree G1 fell along power-law curves as pretraining scaled, and it has open-sourced the weights and code of a Light-O1-Preview text-to-action model.
Image © Qualcomm — The chipmaker is acquiring Boulder-based PickNik, steward of the widely used MoveIt manipulation framework, and says MoveIt 1 and 2 will remain open-source under their existing licenses with third-party hardware still supported. Announced at ROSCon in Toronto, the deal folds motion planning into the Dragonwing robotics platform Qualcomm pitches at everything up to full-size humanoids; the price wasn't disclosed.
Image © Humanoid Guide — A staged studio video shows XPeng's IRON turning toward whoever is speaking using a nine-microphone array plus lip reading, answering in Chinese, English and German from a single multilingual foundation model, and telling three testers apart by face, voiceprint and conversation history, according to the company. The stack runs on three in-house Turing chips with a combined 2,250 TOPS and is aimed at store-guide and car-sales work — though humanoid.guide notes it is a company demonstration with no accuracy figures and no deployment date.
NVIDIA's ROSCon release adds agentic workflows that automate environment setup and repetitive development chores, alongside new Isaac Skills including a FoundationStereo fine-tuning path and a FoundationPose inference library NVIDIA says tracks objects up to 5.5x faster. The GPU-accelerated package set supports ROS 2 Lyrical on Ubuntu 24.04, and Mentee Robotics, Flexiv, Magna and Intrinsic are already using it, according to the company.
Image © Humanoid Guide — The brain-inspired foundation model from Suzhou's JoyIn is designed to connect memory, decision-making, perception and responsive body control in a single system, trained mostly on human video to cut dependence on costly robot-specific datasets. JoyIn reports more than 90% first-attempt success in demonstrations on Unitree and AgiBot robots and points to a ten-minute single-take run as evidence of sustained operation — results humanoid.guide notes remain independently unverified.
Cofounder Gao Yang tells Reuters that Spirit AI's robot brains will reach their “GPT-3.0 milestone” by mid-2027 — hear a spoken request, then attempt a series of reasonable physical actions — with tens of Moz1 wheeled humanoids already working at CATL and JD.com and around 1,000 contractors capturing movement data, according to the company. The date is a forecast rather than a demonstrated benchmark, and Gao concedes fine-motor tasks like unscrewing a bottle cap remain unsolved.
Image © Reward AI — The startup's general-purpose robot policy trains on manipulation data captured by a wearable 7-DOF Omnibody Hand as people simply do their work — no teleoperation, no on-robot data — and picks up a new task from under 30 minutes of demonstrations, according to Reward AI. humanoid.guide notes that no paper, open model or benchmark data is public yet.
Image © Figure — Pretrained on Figure's Index dataset of human behavior, the new policy tackled bed-making, towel-folding and tidying in 30 Bay Area homes it had never seen, scoring 56% across 420 attempts against 9% for a from-scratch baseline — the first zero-shot whole-body generalization at this scope, according to Figure. The company says Helix 2.5 also matched its predecessor on half the adaptation data.
Image © Arm — Arm's new Total Design for Physical AI program brings together more than 80 companies spanning the physical AI stack — from NXP and AWS to Hugging Face, Qwen and Unitree — according to Arm. A companion Robotics Capability Framework grades robots from RL0 reactive systems up to RL5 self-improving ones, meant as a common language for what robots can actually do.
Image © Kinetic Blocks — The training-data startup has published a no-login catalogue at kineticblocks.com/open-data indexing 27 open robotics datasets — around 19,000 hours of robot and first-person human video spanning DROID, Open X-Embodiment and AgiBot World — with each licence spelled out in plain words, according to Kinetic Blocks. The company hosts none of the files itself and says quality and usability checks on the sets will follow.
Image © Humanoid Guide — Meta has released Project SuperDex, a dexterous-manipulation platform that pairs a contact-first physics engine — rigid and soft bodies, tendons and cloth in one solver, with tactile sensing as a native primitive — with robotics authoring tools and a Gymnasium-style reinforcement-learning interface, under an Apache 2.0 license. VR teleoperation is on the roadmap for Q4 2026, and the project's materials show a shape-sorting policy trained entirely in simulation running zero-shot on a real hand — with no benchmark results published yet.
Sequoia is leading a round that would value the two-year-old startup at about $500 million – terms not yet final, according to TechCrunch – just three months after its $60 million Series A. Mecka pays people to record everyday tasks with body sensors and smartphones, turning human motion into training data for humanoid developers.
Standardized tactile data is scaling fast: Tsinghua researchers have pooled more than 3,000 hours across 21 sensor types into one hardware-agnostic model, and Fudan spinout NeoteAI has logged over 30,000 hours of demonstrations, according to IEEE Spectrum. The payoff shows in the tasks VLAs fumble – seating a USB plug, turning a key – where Berkeley's tactile-trained system hit a 65% success rate across 12 tasks, nearly double the best VLA.
Image © Humanoid Guide — RoboGesture generates humanoid gestures from spoken language in real time, tackling limited training data, speech-to-motion mapping and the safety risks of running generated motions on real hardware, according to Galbot. Built with researchers at Tsinghua, Peking University and three other Chinese institutes, the framework is set for ECCV 2026 — with no performance figures or test platform disclosed yet.
Image © Humanoid Guide — Figure plans to commit $3.5 billion in compute with British infrastructure firm Nscale — expandable to $6 billion — for a cluster that could eventually include 100,000 NVIDIA GPUs dedicated to training its Helix models. The first systems are slated for Nscale's Barstow, Texas campus in the second half of 2027, according to the companies.
Image © Humanoid Guide — The Beijing Innovation Center of Humanoid Robotics says its open-sourced RoboMIND dataset — 300,000-plus dual-arm manipulation trajectories spanning more than 700 real-world tasks on 40 robot configurations — has doubled its download count in a month. The figure is self-reported, and as humanoid.guide notes, downloads aren't deployed models or completed training runs.
Image © Humanoid Guide — The Korean firm's Open Humanoid GYM packs simulation, motion development and AI training layers into an open-source program built around its AI Sapiens K1 platform, with students moving full-body behaviors from simulator to physical robot over a one-month program, according to a report cited by humanoid.guide. Key implementation details remain unspecified, and the outlet frames it for now as an educational humanoid test bed.
Image © NVIDIA — The chipmaker has agreed to buy the open AI development platform that also houses LeRobot, the open-source library for robot datasets, models and policies. NVIDIA says Hugging Face will remain open to the entire AI ecosystem — and that NVIDIA compute will not be required to build or deploy on it.
The Sunnyvale startup's first 'live model' generates frame by frame the way an LLM predicts tokens – streaming 4K video at 24 fps, taking prompt changes mid-generation and sustaining hour-scale runs without drift, according to founder Qing Yin. Visko pitches Orbis for robotics training and simulation, opened public access at launch and closed a $10 million pre-seed led by Llama Ventures.
Image © Kinetic Blocks — The Oslo startup's new marketplace lists egocentric video, teleoperation recordings and robot execution data with chain-of-custody documentation and a quality grade, and buyers walk away with a commercial licence at checkout, according to the company. It opens September 1 as a gated beta, with sellers and buyers approved individually.
Skild AI says its flagship S1 foundation model picks up long-horizon tasks like repotting a plant or cooking pancakes from a single human video added to its prompt, and that the same 'omni-bodied' brain runs quadrupeds, static arms and humanoids alike. CEO Deepak Pathak tells The Robot Report that humanoid performance is the next focus – and that robotics still hasn't hit its ChatGPT moment.
Image © Humanoid Guide — Carbon, the control system Sanctuary developed alongside its Phoenix humanoid, handled live-conveyor wire insertion at a global Tier 1 automotive supplier – 313 trials, a 99.5% success rate, 2.54 seconds a cycle, according to the company. Sanctuary pitches the system as Physical AI for currently available industrial robots with off-the-shelf end effectors; for now it's a proof of concept, not a production deployment.
ExlService Holdings has completed its purchase of iMerit, the San Jose firm whose annotation workforce and Ango Hub platform prepare multimodal training, evaluation and reinforcement-learning data for robotics and autonomous mobility. Founder Radha Basu stays on to run the unit inside EXL, which says the deal connects parts of the AI lifecycle that have often been managed separately.
The 25 cm robot waddles, roller-skates and rights itself after falls, and its simulation, SDK and reinforcement-learning stack are open source, so behaviors trained in sim run straight on the hardware, according to Hugging Face. CEO Clem Delangue pitches it as the start of affordable open robots that democratize physical AI and world models, with shipping due before Christmas.
Image © The University of Texas at Austin — The U.S. National Science Foundation is putting $90 million into three new Science and Technology Centers, and the robotics one — the Center for Human and Robot Co-Adaptation at the University of Texas at Austin — gets $30 million over five years to study how people and robots with embodied intelligence adapt to one another as service and assistive robots spread through homes, hospitals and workplaces. Apptronik, Google DeepMind and NVIDIA are among the center's industry partners, according to the university.
Perceptron, founded by two former Meta AI researchers, has released Isaac 0.5, an open-weight vision model built to let industrial robots perceive, reason and act — rejecting what the founders call a false choice between heavyweight generalist models and narrow perception-only systems. It was trained on a million hours of video, from egocentric footage to human demonstrations, according to the company.
Reporting from Actuate — a robotics conference that has tripled in size since 2023, drawing 1,500 attendees according to organizer Foxglove — TechCrunch finds robot hardware outrunning the models that control it, with physical AI sitting in its 'GPT-2 era', as Antioch founder Harry Mellsop puts it. Wayve CEO Alex Kendall reckons manipulation robotics today is roughly where self-driving was five years ago.
Image © NVIDIA — The 78-TOPS module delivers twice the inference performance of its predecessor and, according to NVIDIA, lets robots run frontier LLMs and vision-language models directly on the edge. It is slated for the first half of 2027, with a 15-watt mode that matches the outgoing board's peak output on 40% less power.
The Medal spinout, which trains action models on hundreds of millions of hours of gameplay footage, is raising an oversubscribed round at a $6 billion pre-money valuation — up from $2.3 billion in June — with Valor Equity Partners, Point72 Ventures and Seven Seven Six joining, according to TechCrunch. The new money is earmarked for compute and for extending its general model to robotic embodiments.
Two months after its $400 million June round, the robot-brains developer has added roughly $200 million more led by 8VC, with NVIDIA, Bezos Expeditions and Fei-Fei Li among the backers, according to an SEC filing reported by Axios. TechCrunch says the extension lifts Generalist's Series B to $600 million in total.
The limiting factor in real-world robot deployments is rarely the robot but the workforce that operates, maintains and adapts it, argues HireArt co-founder Christopher Bower in The Robot Report. He describes a new layer of roles sitting between engineering and operations — teleoperators, data capture specialists, QA validators and field technicians.
Image © Humanoid Guide — T-Rex pairs a slower vision-based planner with a fast tactile expert that corrects grips in real time — processing four touch updates for every vision update to catch contact events cameras miss, like a USB connector settling into its port. The open-source release includes 50 hours of robot play data across roughly 5,500 episodes from a 22-degree-of-freedom tactile hand — the largest released tactile dataset to date, according to the collaborators, though no benchmark scores accompany the launch.
Image © Humanoid Guide — Generalist says its new GEN-1.5 system can pick up a physical task from a single demonstration in a few seconds and then generalize what it was shown, crediting pretraining on physical data at scale. The announcement names no benchmarks, success rates or hardware, humanoid.guide notes, so the claim can't be assessed yet.
Image © Humanoid Guide — More than 80% of humanoid robots worldwide are estimated to use NVIDIA's robotics platform — its chips, GR00T foundation models and training environments — according to ABI Research data cited by Chosun. humanoid.guide notes the figure blends two very different dependencies: robots running NVIDIA hardware and those trained with its software.
Generalist told The Robot Report it builds robot manipulation skills from 2–80 hours of human demonstration data per task, then refines them with as little as 4–80 minutes of robot-collected footage. According to the company, the models it has run on Universal Robots and Flexiv cobot arms weren't programmed for error recovery — they learned it from other interactions in its dataset.
Image © Humanoid Guide — The open-source Newton simulation project's v1.5.0 release introduces experimental vectorized joint control with impedance controllers for GPU-scale robot batches, plus multi-world simulation tools and steadier contact modeling. Humanoid benchmark workloads initialize 11–24% faster, according to the release notes.
Humanoids will be judged less by how they move than by how they communicate, argues Treble Technologies co-founder Gunnar Pétur Hauksson — yet the environments robots train in are overwhelmingly visual and, for the most part, silent. The Icelandic firm's simulation platform generates physically accurate acoustic training data with a low sim-to-real gap and is already used by leading technology companies, according to Treble.
Trend Micro subsidiary VicOne has released a free Radeis extension for NVIDIA Isaac Sim that lets developers replay adversarial visual prompts, malicious inputs and sensor manipulation attacks against their own robot models in simulation. The scenarios build on a DEF CON 34 robot-hacking challenge where 16 teams logged 25 successful solves, according to the company.
Image © Humanoid Guide — The Tokyo firm's foundation world model unifies long-horizon reasoning with low-latency control, letting a robot evaluate outcomes through predictive world modeling before acting and keep improving from fleet data, according to Highlanders. Rated 9/10 on humanoid.guide's Brain Score scale, KEPLER comes from the same firm whose roughly 1.75 m humanoid Mitsubishi Motors plans to mass-produce in Kyoto from 2027.
Image © Humanoid Guide — A new video shows the Figure 03 humanoid ascending a ladder with no visible human intervention — 'fully autonomous,' according to founder Brett Adcock, whose Helix system pairs visual perception with whole-body motion control. Figure has not disclosed the control stack behind the clip, and repeatability, speed and whether the setup was staged all remain unverified.
Image © Humanoid Guide — UnifoLM OminiA runs Unitree's G1 as a single system — handling spoken requests, visual scenes, task reasoning and interruption recovery in one framework, with household jobs like laundry sorting and dishwasher loading as the pitch, according to the company. Code and weights stay closed for this version, which is offered via backorder rather than as finished mass-market software.
Founded by leaders of Google DeepMind's applied robotics group, the London- and Sydney-based startup has emerged from stealth with pre-seed backing from Fly Ventures and firstminute capital, building systems that let front-line workers teach robots by demonstration and correct them mid-task. Its first cells are tending 3D printers and disassembling hard drives for materials recovery, and prototyping a new behavior has shrunk from about a day to roughly 10 minutes, according to the company.
NVIDIA's simulation-and-compute stack, Applied Intuition's validation tooling, Scale AI's data factory, Hugging Face's open-source LeRobot layer and Lightwheel's Real2Sim2Real loop are hardening into the load-bearing infrastructure of robot intelligence, according to The Robot Report's analysis. Scale says it now collects more than 1,000 hours of demonstration data a day, after delivering over 150,000 hours of physical AI data to robotics customers in 2025.
Image © Google DeepMind — Google DeepMind's new model family pairs a vision-language-action model that runs a full humanoid — walking, crouching and working 22-DOF five-fingered hands — with an embodied-reasoning model that plans multi-step tasks and coordinates teams of different robots, according to Google DeepMind. An on-device variant adapts to a new robot body in a few hours from under 200 examples, with Apptronik, Boston Dynamics and Agile Robots among the launch partners.
Image © Enigma — The Israeli-Californian lab emerged from stealth with a $71 million seed led by Index Ventures and Ribbit Capital, building AI models and interfaces meant to make any robot as intuitive to command as a car's volume knob, according to the company. Alongside the round, Enigma put what it calls the world's first interactive AI robotics experience online — 100 real robots the public can drive at robots.online.
Image © Humanoid Guide — The 25th IEEE-RAS Humanoids conference lands in Santa Clara this December under the theme 'Humanoids and the Future of Work', putting deployment governance and worker impact alongside actuators and reinforcement learning. A loco-manipulation challenge will benchmark 10–15 academic and commercial teams on standardized factory-style tasks.
Encord has turned a San Leandro facility into a data factory where camera- and sensor-rigged human 'pilots' produce the demonstrations robot models are starving for — an adequate dataset would need something like five times YouTube's video corpus, according to the company. Germany's Zander Labs is testing a stranger input: headsets that read error, intent and surprise from trainers' brains to flag when a model needs the most compute, according to TechCrunch.
Generalist has shown its GEN-1 foundation model driving five-fingered hands, power tools and custom grippers from a single base model — even adapting mid-task when the end effector is swapped, according to the company. GEN-1 was pretrained on more than half a million hours of real interaction data covering roughly 9,000 end-effector variations, Generalist says.
Image © NEURA Robotics — NEURA Robotics and RWTH Aachen University will open a 3,000-square-meter Physical AI training center on the Melaten campus by the end of 2026, where robots collect the real-world interaction data that simulation alone can't deliver. It is one of ten NEURA Gyms planned worldwide — five running across Europe, the US and China by year-end, according to the company, whose CEO David Reger says the biggest bottleneck in Physical AI 'is no longer intelligence, it's experience.'
Vision and joint encoders can't capture contact physics, argues Contactile CEO Heba Khamis — she proposes a 'vision-μ-action' class of models conditioned on real-time friction measurement, shifting world models from statistical pattern-matching toward causal understanding of why grasps succeed or fail.
The Nvidia-backed startup founded by ex-DeepMind researchers will use Elite Robots' industrial cobots for real-world data collection and validation of its GEN-1 foundation model — a pairing that has already logged 1,800+ block stackings with zero interventions at a 99% task success rate, according to Elite Robots.
Image © Tencent — At WAIC in Shanghai, Tencent rolled out a full-stack embodied intelligence offering — pairing the Hy-Embodied-VLA-0.5 action model with the VLM-1.0 perception and RxBrain-1.0 reasoning models under a new TairosAgent framework, according to Tencent. The Tairos platform, first launched in 2025, now ships open-source tools and cloud services for robot developers.
Teleop datasets are over 100,000 times smaller than what trains today's language models, Nikita Rudin argues — and reinforcement learning, not an endless stream of human demonstrations, is the way out.
Image © Sunday Robotics — The Memo maker's new foundation model folded clothes in homes it had never seen at a 99.1% success rate — 778 of 785 attempts, zero-shot with no per-home fine-tuning, according to Sunday Robotics, though the figures are the company's own evaluation. ACT-2 was pre-trained on manipulation data from $200 skill-capture gloves shipped to more than 2,000 contributors.
Image © Agility Robotics — The Digit maker's new Fremont center will focus on training, testing and advancing the AI that teaches its humanoid new skills, according to Agility — with nearly 200 AI and field-operations hires planned and more than $300 million in multi-year Digit orders behind the expansion.
Image © LimX Dynamics — The freshly funded Shenzhen maker's three-layer 'agentic OS' pairs a world model and closed-loop reasoning with VLA skills — enough, according to LimX, for China's first fully autonomous long-horizon home run by a full-size humanoid.
Image © Humanoid Guide — NVIDIA's AI director told RAISE Week in Paris that humanoid prices have fallen 98% in two decades and open GR00T downloads have passed 10 million, according to Fan — with 'world action models' that predict future frames as the next paradigm.
Image © 1X Technologies — The NEO maker's new research organization, headed by ex-Luma researcher Samarth Sinha, will pretrain on web media, human video, simulation and robot data — world models, says CEO Bernt Børnich, will free humanoids from narrow tasks in static environments.