A weekly digest of Physical AI — humanoids, robot foundation models, world models, sim, data — curated by people.

Learning Safe Humanoid Navigation from Reduced Order Models (paper · 15 minute read)

A single-stage RL navigation pipeline struggles on multi-story terrain, so the authors split the problem: a policy on reduced order dynamics with full 3D LiDAR observations learns multi-story navigation, then kickstarts a full-order policy with a frozen locomotion policy in the loop. A Poisson safety filter recovers safety around out-of-distribution obstacles. On a Unitree G1, RoM-Nav performs mapless multi-floor navigation over 10m of vertical displacement and over 100m of path length.

China's Ubtech Opens Plant Making a Humanoid Robot Every 10 Minutes (3 minute read)

UBTECH has started operations at a humanoid robot plant in Liuzhou, Guangxi, with annual capacity of 10,000 units, according to Global Times. Built with Siemens Digital Industries Software and covering 14,000 square meters, the plant mainly produces the Walker S and Cruzr series, with Cruzr Y1 and Cruzr S2 humanoids handling materials and more than four hours of full-unit testing for each finished robot.

Agility Robotics says its new Digit 5 robot can work next to people without safety fences (1 minute read)

Agility Robotics' Digit 5 lifts up to 22.7 kg, 40 percent more than its predecessor, and its battery charges in 9 minutes for 90 minutes of runtime, which the company says allows more than 20 hours of work per day. The humanoid spots people and, according to the company, stops or steps aside. Agility points to orders worth more than $300 million, with first deliveries starting in early 2027.

China scales up embodied AI's reach (2 minute read)

Agibot says its 15,000th robot rolled off its production line in June. Eight of its G2 robots ran around the clock for six days alongside a mass-production line at a Longcheer Technology tablet factory in Nanchang, China Daily reports. Galbot's S1 has entered CATL battery production lines and can carry a combined 50 kilograms with its two arms for up to eight hours, according to CATL.

Helix 2.5: Zero-Shot 30-Home Generalization (8 minute read)

Figure has pretrained Helix 2.5 on Index, its dataset of human behavior, and run three whole-body behaviors — tidying living rooms, folding towels and making beds — in 30 Bay Area homes with no data collected in any of them. The company reports that Index pretraining alone lifts zero-shot success from 9% to 56%, and that Helix 2.5 matches a Helix 02 policy while using half the task-specific data.

Does Scaling Web-Video Pre-Training Help Real Robots Do Real Work? (13 minute read)

Larger pre-trained video models consistently yield better robot policies, and the trend still holds at the largest size tested, the authors report from a scaling study of Direct Video-Action models pre-trained on web video. More pre-training compute improves policies at every amount of robot data tested, with the largest gains when robot data is scarce. The findings rest on over 200 hours of real-robot evaluation on an industrial bearing-unpacking task.

Isaac Lab 3.0 Early Access puts one task API across physics and rendering backends (tool · 8 minute read)

Isaac Lab 3.0 Early Access brings one task API across multiple physics, rendering and visualization backends, so supported tasks can switch between them without rewriting the environment. Many workflows now run kit-less, without installing or launching Isaac Sim, and a new isaaclab command replaces library-specific training and playback scripts. Quaternions change from WXYZ to XYZW, and General Availability is targeted toward the end of October 2026.

EgoSteer-RealWorld: 192 hours of teleoperated bimanual manipulation across 193 tasks (dataset · 1 minute read)

EgoSteer-RealWorld, the real-robot dataset used in EgoSteer, holds 54,454 teleoperated episodes (192 hours, 20.75 M frames) of bimanual dexterous manipulation across 193 tasks on a RealMan dual-arm robot, according to its Hugging Face dataset card.

FoldNet++: a Large-Scale Synthetic Dataset for Robotic T-Shirt Folding and Unfolding (paper · 15 minute read)

FoldNet++ spans 6 robotic embodiments, 1K T-shirts, 1K environmental assets and 120K episodes with rich annotations. It generates physically simulatable T-shirts with annotated semantic keypoints, from which a unified rule-based framework produces demonstrations for each embodiment. The authors report that visuomotor policies trained solely on this synthetic data achieve over 90% end-to-end task success when deployed directly to unseen real-world environments and previously unseen T-shirts.

D-Robotics Closes $400M in Series C Funding to Boom the Robotics Industry (2 minute read)

D-Robotics has closed a US$400 million Series C, with strategic investment from an unnamed global internet company and follow-on from existing shareholders. The round funds its Sunrise chip portfolio and a software platform. The company says cumulative Sunrise shipments passed 8 million units in the first half of 2026, and that its S600 embodied AI chip, launched in November 2025, has been adopted by more than 20 customers in six months.

RoboHarm: Do Frontier Robot Policies Refuse Unsafe Instructions? (4 minute read)

SoftBank agrees to acquire Robotics and AI Institute (1 minute read)

Robotic lab sets up and runs optics experiments on demand (5 minute read)

Isaac Sim 7.0.0 Alpha 1 is tagged on GitHub (tool · 1 minute read)

PreDE predicts offline whether quantization will degrade a world action model (paper · 15 minute read)

ForwardDLO: Model-Based Bimanual Shape Matching of Unconstrained Deformable Linear Objects (paper · 15 minute read)

Agility Foundations: Functional Safety (video · 3 min)

BlackPanther2 quadruped runs at 13.2 m/s on a treadmill and 11.65 m/s outdoors, authors report (paper · 15 minute read)