Termsslug: vla

VLA

视觉-语言-动作模型

Model paradigm jointly modeling visual perception, language instructions and action outputs, common in robotics and game agents.

Curated Intelligence

31 items · Page 1/4
量子位官网T2·Models
78 pts
miHoYo Unveils 8B-Parameter Gamer VLA Model That Plays Star Rail Autonomously

miHoYo's Honkai AI research team released Gamer, a general-purpose vision-language-action (VLA) foundation model built on Qwen3-VL-8B that outputs keyboard and mouse actions from screen pixels and task instructions alone, without reading game internals. In a roughly 50-hour long-horizon evaluation across five main story chapters of Honkai: Star Rail, it ran autonomously for over 95% of the time, unifying cross-game real-time control with long-horizon memory and planning in a single VLA framework.

InfoQ 中文T2·Models
78 pts
Google Releases Gemini 4 Argon for Robot Control and Simulation

Google released Gemini 4 Argon, positioned as a general-purpose physical reasoning model for robot control, Real2Sim scene construction, and iterative policy optimization rather than a single end-effector policy. Google reports that after training in a MuJoCo twin environment, Argon guided Boston Dynamics' Spot through a foot-pedal task, and that domain-randomized MuJoCo scenes synthesized from a single RGB photo lifted a vision-language-action policy's real-world success rate from 56% to 75%; it also scored 87% on the ERQA embodied reasoning benchmark using pure reasoning. The model is available to Google AI Ultra subscribers and via the paid Gemini API, though the demos rely on lengthy task-specific prompts and verifiers, and failure rates were not disclosed.

量子位官网T2·Agents
58 pts
Gen-Z Team Wins Embodied AI Challenge with Self-Collected Data

Ant Group's Lingbo held its first embodied large-model challenge, requiring 15 finalist teams to complete data collection, model tuning and real-robot deployment within 48 hours, with one model handling fruit picking, cosmetics shelving and dual-arm sweeping. The TARS team, averaging 25 years old, discarded the official 210 demonstrations that grasped the dustpan by its handle and re-collected its own data gripping the rear edge, overtaking the runner-up at the last minute with 82 points and an 80,000 RMB prize. The case shows that in real-robot VLA tasks, data collection strategy and contact-point choice can matter more than model tuning.

量子位 (微信公众号)T2·Models
72 pts
Zhejiang University's VLA-Delta Teaches Robots Explicit Physical State Change

Zhejiang University's ACES Lab OmniAI team introduces VLA-Delta, which turns task progress, spatial displacement and contact changes into explicit pretraining supervision so the representation behind action generation learns both task semantics and visual motion. Built on a 2B vision-language backbone, it reaches 98.4% average success on LIBERO and 68.0% on five real dual-arm robot tasks.

The DecoderT2·Open Source
78 pts
Runway Open-Sources Praxis-1 Robot Control Model

Runway has released Praxis-1, its robot control model, with open weights and source code under Apache 2.0. Trained on video and robot data, it maps camera images, text instructions and position data to robot motions, and can be adapted to new robot bodies via additional training. It still confuses similar-looking objects and cannot yet reliably plan complex action sequences on its own.

机器之心 (微信公众号)T2·Models
78 pts
Stanford Fixes Robot Instruction Failures by Editing 65 Embedding Rows

A Stanford-led team proposes debugging robot policies by directly editing 65 rows of a vision-language model's token embedding layer, with no gradient updates or new demonstration data. On 335 held-out instructions containing the edited tokens, success rate rose from 74.3% to 87.5%, while outcomes on 430 other episodes stayed unchanged. The paper, code and all 16 experiment logs are public.

IT之家 · 智能时代T2·Industry
62 pts
Beijing Unveils Three-Year Embodied AI Robot Action Plan

Beijing's economy and information technology bureau issued a three-year action plan for embodied AI robots, pledging one-on-one IPO support for qualifying firms and cultivating leading players. The plan targets at least 200 tech SMEs and specialized firms across complete robots, key components, embodied models and data collection, while backing VLA, world models and reinforcement learning routes.

IT之家 · 智能时代T2·Models
88 pts
Google Unveils Gemini Robotics 2.5 VLA Model for Real-World Robots

Google has launched Gemini Robotics 2.5, its most capable vision-language-action model to date, delivering gains in dexterity, reliability and safety across more than 200 tested tasks. In a real-world proof of concept on a Mercedes-Benz production line, it drove Apptronik's Apollo 2 humanoid to pick and pack auto parts with a 96% success rate, improving success by 30-100% and speed by 20-30% over prior baselines. The model is now opening to Google's robotics partners.

The DecoderT2·Agents
88 pts
Gemini Robotics 2.5 Hits 96 Percent Success on Mercedes Assembly Line

Google DeepMind unveiled Gemini Robotics 2.5, a vision-language-action model it says outperforms its own prior baselines by 30 to 100 percent depending on task type. On active Mercedes-Benz production lines in Hungary and Germany, Apptronik's humanoid Apollo 2 sorted metal parts with a reported 96 percent success rate. DeepMind also uses Gemini 4 Argon as a simulation engine for training and open-sourced the ASIMOV-Jailbreak benchmark to test robotics model safety.

Ars Technica AIT2·Models
78 pts
DeepMind Unveils Gemini Robotics 2.5 and Open Safety Benchmark

Google DeepMind has released Gemini Robotics 2.5, a vision-language-action model with substantially improved dexterous manipulation that has already performed real work in Mercedes-Benz factories and enables Apptronik's Apollo 2 humanoid to tie a bow with five-fingered hands. The team also introduced ASIMOV-jailbreak, an open safety benchmark aimed at evaluating robots operating alongside humans.