VLA

Headlines · 5

1. Oct 26
   Researchers Explore GPT-6 Astra for Robot Control

   36Kr, citing a CITIC Securities research report, says general-purpose models are extending beyond high-level robot task planning toward action generation and execution control. The report says NVIDIA, Stanford and other research institutions have been exploring GPT-6 Astra for robot control, using its reported visual understanding, spatial reasoning and task-planning abilities to reduce reliance on specialized policy training. The supplied reports describe research exploration rather than a confirmed production deployment.

   36氪快讯 · 🔥 0
2. Oct 23
   2027 Tank 500 launches with VLA-powered driver assistance

   DoNews reported that the 2027 Tank 500 launched at 335,000–375,000 yuan, with 296 software, hardware and safety-redundancy upgrades. The vehicle uses the Coffee Pilot 4 driver-assistance system, which Tank says combines a VLA model, a Thor-U chip and 27 sensors; launch benefits include lifetime free use of the VLA driving assistance system. Tank also introduced the Danxia Red exterior color and a revised cabin featuring dual 15.6-inch displays and 22 speakers.

   DoNews · 🔥 0
3. Oct 22
   XPeng says second-gen VLA model runs on-device in IRON robot, re-reasoning every 0.2 seconds

   XPeng announced that its second-generation VLA large model has been upgraded with stronger temporal understanding and future-behavior prediction, and that this capability has now landed on-device in its IRON humanoid robot, according to IT之家. The company says IRON re-reasons from the image in front of it every 0.2 seconds and predicts motion changes within the next 6 seconds, instead of recognizing a gesture once and then executing a fixed sequence. XPeng claims the robot can adapt mid-interaction during handshakes, fist bumps and high-fives — adjusting when a hand suddenly closes into a fist or when the target position of a high-five changes — closing a real-time loop between what it sees and how its body moves; all of these are vendor claims. IT之家 noted that XPeng unveiled the new-generation IRON at a tech day on November 5, 2025, with a physical-world model, three Turing AI chips and a VLT+VLA+VLM combination, and that XPeng plans mass production by end-2026 with deliveries in the first half of 2027 starting in stores and shopping malls.

   IT之家 AI · 🔥 0
4. Oct 19
   Dyna Robotics unveils Dyna-2 and Taku robot, says model route shifted from VLA to world-action model

   Dyna Robotics has released Dyna-2 and a new robot, Taku, which performs an end-to-end laundry workflow — loading dirty clothes into a washing machine, operating it, moving clothes to a dryer, then folding and shelving them — as shown to Silicon Valley 101 in an interview with co-founder Jason Ma, republished by TMTPost and 创业邦. Ma said Dyna moved its model route from VLA (vision-language-action) to a world-action model (WAM) and expanded pretraining data to 1 million hours of first-person human video, and that the model predicts robot actions unseen in pretraining increasingly well as human video data grows. Those capability claims come from the co-founder rather than independent evaluation, and Ma said real deployment is the only evaluation for robots, arguing that failures to deploy stem from models and systems rather than from lacking legs or hands.

   钛媒体 · 🔥 0
5. Oct 12
   BAAI's Wang Pengwei on RoboBrain Orca: world models vs VLA for embodied generalization

   Wang Pengwei, who leads the embodied model research center at China's BAAI, presented the RoboBrain Orca world model at the 2026 BAAI conference. He argues that today's four families of world models — language-, pixel-, 3D-structure- and visual-representation-centric approaches such as LLMs, video generators, World Labs' Marble and JEPA — each capture only a partial world and remain passively task-driven; Orca instead uses Next State Prediction, reasoning forward and backward about state transitions, and separates "unconscious" video learning from "conscious" event-level state-transition learning. He says Orca's pretraining used no action data, drawing on about 125,000 hours of internet video, roughly 130 million expert-model-labeled event-level state-transition samples and causal and general VQA data, and that BAAI is building a benchmark for state-transition world models because none exists; he expects some approach to break the 0% generalization barrier within a year, though 90%-plus would be needed for real usability.

   钛媒体 · 🔥 0

Experience and discussion from the community

Share my VLA experienceAsk about VLA