Gemini Robotics 2.5 hits 96% kitting parts on Mercedes lines
DeepMind's new robot model lifts task success 30–100%, and Apptronik's Apollo 2 used it to sort car parts on live Mercedes lines.
Embodied AI and hardware. 11 stories.
Robotics is where AI meets the physical world: manipulators learning from demonstration, humanoids moving into pilot deployments, and foundation models trained on robot data. Progress here is slower and harder to fake than in software, so the interesting stories are about what works outside a demo.
This section covers research results, hardware launches and real deployments, with a bias toward evidence of reliability over staged videos.
DeepMind's new robot model lifts task success 30–100%, and Apptronik's Apollo 2 used it to sort car parts on live Mercedes lines.
A third-party Unitree G1 overheated in China's Mount Tai robot race and was stretchered off. A Unitree A2 robot dog won the legged race.
China's BG-5 robotic arowana became a spy-gadget meme at CIFTIS. The joke is half right — it's built for port recon.
Figure's Helix 2.5 did chores in 30 Bay Area homes with no retraining. It succeeded 56% of the time, against 9% for a model trained from scratch.
UBTech's Liuzhou plant now rolls out a humanoid every 10 minutes — 10,000 a year, with robots doing much of the assembly.
A user hooked OpenAI's week-old Astra to a robot arm and a camera. It learned the controls and painted the Golden Gate Bridge, better on every try.
EngineAI's T800 humanoid strolled San Francisco streets — teleoperated by REK Robots, the fight-club startup that imported the first one to the US.
Neura's 4NE1 folded mid-demo in Taipei. Qualcomm says it was a safety feature, not a crash. The internet disagreed.
DeepMind's new VLA model controls humanoids head-to-toe, adapts to any two-armed robot in hours, and coordinates robot teams.
Satyress's Threehalves bolts a humanoid torso onto a four-legged base — and there's no AI in it. Just a human on a joystick.
Google DeepMind's new robotics model lets robots inspect instruments, detect task success, and refuse unsafe actions — now live via API.
Feed, daily deep-dive and bytes — readable offline, with push alerts for the topics you follow.