Robotics and Physical AI

Oct 27

Oct 27

NVIDIA published two ONNX exports of its Metatron G1 walking and recovery policy for browser demos. Here is what the model card says, what you can and cannot do with the files, and where it fits in NVIDIA humanoid work.

Oct 27

A Purdue IDEAS Lab video of a drone trailing a runner through dense forest without GPS or cloud is spreading. DroneXL spotted open-source University of Hong Kong code in the telemetry. We separate the demo, the credit dispute, the Army grant and what is still unpublished.

Oct 23

NAVER AI Lab and university co-authors posted LaLa-World on October 20, 2026. It describes every robot action in words so a single JEPA-style world model can learn from simulators, robot logs, images and video. The authors report 6x lower training cost and 13.9x faster planning than a matched video-generation baseline.

Oct 23

On October 23, 2026 NVIDIA published a step-by-step workflow where AI agents turn one prompt into a simulated obstacle course for the Microduck robot, and Isaac Lab trains a walking policy on it. Here is what the agents do, what runs locally versus in the cloud, and what the post leaves unmeasured.

Oct 23

Tencent Hunyuan published Hy-Nomas on Hugging Face on October 23, 2026: a segment-autoregressive video world model you steer with text and camera paths, plus Hy-Nomas-Bench with 280 test cases. Here is what it does, what it needs and what is still missing.

Oct 22

NVIDIA open-sourced Flora, the data engine behind GR00T-2 training: a segmenter, a captioner and a reward model for robot and human manipulation video. The weights are real; the 125x cost claim is a researcher post, not a model-card number.

Oct 22

A Berkeley team including Pieter Abbeel and Koushil Sreenath reports SuperMimic, which trains one world model and one policy together and uses analytic gradients. In 3 hours on one GPU it passes the return of a tracker trained for 7 days on 128 GPUs.

Oct 22

Tesla says the first production-quality Optimus robots are coming off a Fremont line designed for one million units a year. Musk also cautioned that the ramp will be bumpy. explainx.ai separates the confirmed start of production from the unverified volume, price and delivery claims.

Oct 22

On October 21, 2026, Unitree posted a video of the H2-X "Kung Fu Legend Guardian" and advertised about eight hours of endurance. The on-screen text says standing endurance. Here is what the video shows, what Unitree has not said, and how to read a staged humanoid demo.

Oct 22

Former OpenAI science lead Kevin Weil has launched Warp Sciences, a startup that wants human chemists, wearing cameras and sensor gloves, to train robots to synthesize and test drug candidates around the clock. The pitch is automation of wet-lab work now outsourced to contractors in China. Here is what has been reported, what the data strategy is, and what remains to be shown.

Oct 21

Researchers at Binghamton University demonstrated that a hidden trigger word planted in an LLM can make a robot execute a harmful plan, and built PROTEA, a safety layer that checks each planned step against the environment. Here is what the work shows and what it does not.

Oct 21

Skywork AI released Matrix-Game 4.0 on October 15, 2026 and published the paper on Hugging Face this week. Instead of asking a video model to invent game rules, it splits the job: a lightweight engine keeps state and physics, and a distilled video model paints the pixels. Here is what that changes, the numbers, and how to run it.

Oct 20

Dream Labs came out of stealth on October 19, 2026 with a $90M seed round and Dream-1, a World Action Model it says can drive bimanual robots it was never trained on. Here is how it works, what it costs per action, and which claims still rest on the company alone.

Oct 20

A new benchmark, RoboValue, evaluates the models that score how well a robot is doing a task. It covers four capabilities across 15 simulated and 20 real-world manipulation tasks, and finds persistent errors in instruction grounding, progress tracking and recovery. Code and training data are public.

Oct 20

On October 19, 2026, Skyfall AI unveiled SKY-001, which it calls the first world model for engineering. It turns an image into a tree of CAD parts and a step-by-step build plan, and it turns a building outline into a floor plan. The company's own benchmarks say it matches GPT-6 Astra on quality at a fraction of the time and cost. This post separates the headline from the table.

Oct 20

Oct 17

An AI data lab called HUMXN says homeowners in Minneapolis, Chicago and Miami can get plumbing, HVAC, electrical and cleaning work free if technicians wear capture gear that records the job for robotics training. Here is what is stated, what is not, and why physical-AI labs are suddenly paying for real-world data.

Oct 16

Dynamic Creatures, Boston Dynamics official entertainment and hospitality partner, released a video of Hugo, a character robot said to run a Vision-Language-Expression model. The Boston Globe reports it is currently teleoperated. Here is what is claimed and what is verified.

Oct 16

Hao AI Lab at UCSD released ORA, an open robot-use agent harness that connects open-weight language models to robots with no VLA, robot training or task memory. With DeepSeek-V4.1-Flash it reports 69.2% on RoboTwin and 81.4% on LIBERO-Pro. Here is how it works, the cost, and the caveats the authors list.

Oct 16

A new benchmark built on LIBERO asks robot policies to follow user preferences, such as moving slowly or choosing dairy, without changing the task goal. The strongest zero-shot model completes 97.6 percent of plain tasks but satisfies motion and skill preferences only about a quarter of the time. Here are the suites, the numbers and the open questions.

Oct 14

On October 13, 2026, a five-person team at SberRoboticsCenter posted GreenWAM, a world foundation model that predicts how a robot's actions change a scene and also picks those actions, across 20 robot families. Four benchmark checkpoints are on Hugging Face. This post explains the recipe, the numbers and what the authors have not shown yet.

Oct 13

Nature Will introduced Natura-1 on October 13, 2026: a 50-billion-parameter mixture-of-experts model pretrained on one million hours of egocentric human video that follows language instructions zero-shot and drives 20-DoF dexterous hands in a 100 ms control loop. Here is what the company reports, how its benchmark was built, and what is still unverified.

Oct 13

On October 13, 2026 Physical Intelligence announced the Pi API beta: upload demonstrations, launch fine-tuning on compute it manages, and run the resulting model through a hosted gateway. Here is how the workflow works, who already used it, and the open questions about price, latency and access.

Oct 9

Odyssey launched Odyssey-3 on October 8, 2026, claiming a new state of the art on Physics-IQ Verified and first place in three of four WorldMark categories. We walk through the leaderboard, the cost-per-video math, the robot, humanoid and driving demos, and the caveats on best-of-N scoring and who ran which evaluation.

Oct 6

Reka AI released a research preview of Rho-1 on October 5, 2026: a 19-billion-parameter omni-reasoning model that treats text, images, video and robot actions as tokens in one shared context, with no tool calls or external models. It was trained on 320 H100 GPUs in about three months. There are no published benchmarks yet. Here is what the design means and what to check.

Oct 1

On September 30, 2026 Figure published F.02 Decommission: robots trained in San Jose leaped into a 75-ton electric arc furnace in Imatra, Finland. The point was IP destruction, not a new Helix SKU. US and Mexican foundries refused lithium-ion batteries; the film is not generated footage.

Sep 25

Menlo Research released the full reinforcement-learning training pipeline behind Asimov 1 — its $499-deposit, developer-oriented humanoid robot — covering the PPO and Adversarial Motion Priors code, the Isaac Lab simulation setup, and the sim-to-real deployment path that took the robot from zero to walking. It's a genuinely open humanoid stack, not just an open hardware BOM. Here's what's in the repo and what it's actually good for.

Sep 25

On September 21, 2026, Odyssey shipped a playable research preview of Agora-2, a multi-agent world model that supports up to 20 humans and agents in one shared simulation with streaming pixels and explicit shared state — five times Agora-1 capacity and multiple environments. AWS demoed it live with Matt Wood. Here is how it differs from single-player world models and what robotics and agent trainers should watch.

Sep 21

A viral video from a Friday-night event in San Francisco shows a human fighter taking real kicks from a six-foot EngineAI T800 and struggling to put down a remote-controlled Unitree H1. The clip is being framed as the birth of a new entertainment genre — but the robots were teleoperated, not autonomous, which is the detail that determines whether this scales safely or becomes a regulatory flashpoint for humanoid robotics.

Sep 21

Three years after leaving Meta's chief AI scientist role partly over this disagreement, Yann LeCun reposted his own 2023 claim that autoregressive LLMs will not lead to human-level AI, and defended it point by point, citing JEPA, missing Level-4/5 self-driving cars, and Piaget. Jürgen Schmidhuber immediately disputed the JEPA originality claim underneath it, reviving a priority dispute he has pressed since 2025.

Sep 18

On September 18, 2026, Figure released Helix 2.5 and rented 30 Bay Area homes to prove a specific, falsifiable claim: a single foundation model, pretrained on human video, can walk into a home it has never seen and tidy a room, fold towels, or make a bed with no fine-tuning in that home. The headline ablation is the real story — Index pretraining alone took zero-shot success from 9% to 56%.

Sep 17

Brett Adcock posted eleven words and 444,800 people looked. The post itself says nothing, but three verifiable things published in the weeks before it narrow the space of what Figure can plausibly be showing, and they all point in the same direction.

Sep 16

A biomimetic robotic fish shown at a Beijing trade fair went viral on Reddit and X as a "surveillance" device. The company's own specs describe something more specific and more interesting to AI builders: autonomous obstacle avoidance and multi-robot formation control. Here's what's confirmed, what's headline inflation, and why the coordination stack is the part worth studying.

Sep 16

Odyssey's newest foundation world model, Odyssey-3, claims a material advance in physical accuracy over its predecessors — and unlike most world models built for a single domain, it's positioned as one model powering robots, humanoids, cars, drones, and video games at once. Here's what's confirmed, what isn't, and how it fits the broader 2026 world-model race.

Sep 15

OpenArm is a fully open-source, 7-degree-of-freedom humanoid arm from Enactic built for physical AI research — teleoperation, imitation learning, and contact-rich manipulation — with a complete stack of hardware, ROS2, Isaac Lab, MuJoCo, and dataset repos. A full bimanual system starts at $6,500.

Sep 15

Reward AI introduced OM-1, an in-house policy that learns manipulation from people wearing a multimodal hand device and runs across arms and humanoids. The important caveat: its controller still uses simulation, and the company has not released a paper, weights, dataset, or independent benchmark.

Sep 15

A viral post claimed China opened the world's first mass-production plant making a humanoid robot every 10 minutes. That's real — it's UBTECH's new factory in Liuzhou, Guangxi — but "one every 10 minutes" describes the line's designed pace, not its planned annual output, which is a fraction of what that rate implies if sustained year-round.

Sep 13

Sep 8

Unitree reportedly showed off UnifoLM-X2 on September 8, 2026 — described as the first humanoid combat demo built on a world-model approach, rather than pre-scripted choreography. Here's what that distinction means for robot control and how it connects to the world-model race explainx.ai has tracked all year.

Sep 8

Sep 7

Sep 7

Sep 5

Sep 4

Sep 2

Sep 1

Sep 1

Aug 31

Aug 27

Aug 27

Aug 26

Aug 26

Aug 24

Aug 23

Aug 23

Aug 22

Aug 22

Aug 18

Aug 17

Aug 13

Aug 11

Aug 11

Aug 8

Aug 5

Jul 29

Jul 29

Jul 25

Jul 24

Jul 23

Jul 22

Jul 21

Jul 19

Jul 14

Jul 10

Jul 9

Jul 5

Jun 23

Jun 23

Jun 22

Jun 20

May 30

May 19

May 9

May 6