Skip to content

#具身智能

2 today

Jun 2Tuesday

Latent Space

[AINews] NVIDIA Cosmos 3, Nemotron 3 Ultra, and RTX Spark

NVIDIA released Cosmos 3 and Nemotron 3 Ultra; Cosmos 3 uses a Mixture-of-Transformers design with 16B Nano and 64B Super variants, while Nemotron 3 Ultra is described as a 550B-A55B open-weight model.

Why it matters: Nvidia used Computex to put out a world model, a large model and local inference hardware, with Cosmos 3's hybrid Transformer and 16B/64B sizes as the concrete anchors.

Jun 1Monday

AI HOT (Curated Pool)

NVIDIA Open-Sources Cosmos 3, Its First Generalist Model for Physical AI

NVIDIA open-sourced Cosmos 3 at GTC Taipei, releasing two variants, Super 32B and Nano 8B, with model weights, code, and datasets made available.

Why it matters: NVIDIA open-sourced all of Cosmos 3's weights, code and datasets in 32B and 8B versions, giving robotics work a ready physical-world understanding and action-generation model.

Synced · WeChat

OpenAI recruits for robotics team led by Sora creator Aditya Ramesh

OpenAI has listed more than a dozen San Francisco robotics roles for OpenAI Robotics, a team that evolved from Aditya Ramesh’s Worldsim work, with the actuator design engineer role offering $342,000 to $445,000 in base cash pay plus PPU incentives.

Why it matters: OpenAI is hiring more than ten robotics roles in San Francisco under Sora lead Aditya Ramesh, with actuator design engineers paid $342,000 to $445,000 in base cash.

Synced · WeChat

World models get a “save state”: VAST releases Project Eden

VAST released Project Eden, a three-layer world-model architecture that separates persistent state evolution from visual rendering, and disclosed nearly $200 million across its A+ and A++ funding rounds.

Why it matters: VAST's Project Eden splits a scene's persistent state from its rendering, so a generated 3D world can be saved and reopened with changes intact.

QbitAI · WeChat

How Cloud Models Reach the Physical World: CMG Lion Rock AI Lab Uses LiOS for Embodied AI

CMG Lion Rock AI Lab released the LiOS edge-cloud architecture for embodied robotics, reporting about 30 ms one-way latency from local camera to cloud GPU memory in cross-machine tests, and open-sourced the low-latency video transmission module plus the LeFold laundry-folding dataset.

Why it matters: China Merchants Group's Lion Rock Lab open-sourced the LiOS edge-cloud stack, measuring about 30 ms one-way latency from a local camera to cloud GPU memory, plus a low-latency video module and the LeFold clothes-folding dataset.

QbitAI · WeChat

VAST Raises Nearly $200M and Discloses Its Project Eden World Model Roadmap

VAST raised nearly $200 million in A+ and A++ rounds and disclosed Project Eden, a world model architecture that separates state evolution from visual rendering through a structured state layer, a conditional interface layer, and a generative rendering layer.

Why it matters: VAST published its Project Eden world model architecture alongside a nearly $200 million raise, splitting state inference from rendering into a state layer, a conditioning interface and a generative renderer.

AI HOT (Curated Pool)

Introducing Cosmos Coalition

Runway joined Cosmos Coalition as a founding member and will co-develop the first open world-model foundation model for physical AI with NVIDIA.

Why it matters: Runway joined the Cosmos Coalition as a founding member to build open physical AI world models with NVIDIA and other labs, and its jointly developed foundation model is the alliance's first project.

AI HOT (Curated Pool)

Cosmos 3 Released: First Open Physical AI Generalist Model

NVIDIA released Cosmos 3 as an open physical AI generalist model with native visual reasoning, world generation, and action generation, offering two variants: Super at 32B parameters and Nano at 8B parameters.

Why it matters: Nvidia open-sourced Cosmos 3 for physical AI in two sizes, a 32B Super and an 8B Nano, folding visual reasoning, 3D world-state generation and action output into one model.

AI HOT (Curated Pool)

MWC26 Shanghai to Host First Humanoid Robot Penalty Shootout With Unitree and 7 Other Teams

MWC26 Shanghai will host a humanoid robot penalty shootout in June 2026, with eight Chinese embodied intelligence teams competing under rules that require autonomous play without human control or preset scripts.

Why it matters: Eight teams, including Unitree, must let their robots find the ball and keeper on their own, with no remote control or preset scripts.

AI HOT (Curated Pool)

OpenAI enters robotics and starts hiring

OpenAI formed the OpenAI Robotics team and is hiring full-stack hardware, systems, and ML engineers; Aditya Ramesh leads the project, with a near-term focus on supporting skilled workers, while the post does not disclose hiring scale.

Why it matters: OpenAI formed a robotics team led by Aditya Ramesh and is hiring hardware, systems and ML engineers, aiming at hardware-software co-design starting with skilled-trade tasks.

May 31Sunday

Xinzhiyuan · WeChat

Fudan-Linked Team Releases STI-WM Spatiotemporally Integrated World Model

MouShen Intelligence released STI-WM, a spatiotemporally integrated world-action model for robotics, claiming support for RGB, point-cloud, and proprioceptive inputs, hundred-second task planning, and disclosing five funding rounds in six months plus a RMB 300 million Pre-A round.

Why it matters: A Fudan-linked team's world action model fuses RGB, point clouds and robot proprioception, planning tasks over a hundred seconds long.

QbitAI · WeChat

Robot-Native World Action Model Debuts With Spatiotemporal Architecture From Fudan-Linked Team

Moushen Intelligence released STI-WM, a spatiotemporally integrated world action model for robotics, with RGB, depth point cloud, and proprioceptive inputs; the post says it supports hundred-second-scale long-horizon task rollout and closed-loop replanning, but does not disclose benchmark scores or deployment costs.

Why it matters: Muchen Intelligence released STI-WM, feeding RGB, depth point clouds and robot joint states into one spatiotemporal architecture, supporting long-task rollouts over a hundred seconds and replanning after errors.

AI HOT (Curated Pool)

Tesla FSD completes a 6,000 km zero-intervention autonomous drive across Canada

Tesla FSD V14.3.3 completed a 6,051 km zero-intervention drive from Vancouver to Halifax in 4 days and 21 hours, with the system handling lane changes, complex road conditions, and parking without disengagements or human corrections.

Why it matters: A 6,051 km run with no driver takeover shows what L2 assisted driving can already handle on a long trip.

May 30Saturday

Financial Times · Technology

UK military looks at allowing lethal strikes without human approval

The FT headline says the UK military is examining lethal strikes without human approval, but the accessible body is a subscription page and does not disclose the weapon types, approval mechanism, legal conditions, or deployment timeline.

Why it matters: The UK military is studying whether weapons systems could carry out lethal strikes without human approval, touching the long-running dispute over autonomous weapons and human control.

Synced · WeChat

NVIDIA and Tsinghua Team's Gamma-World Tops Hugging Face Daily Chart

NVIDIA, Tsinghua, University of Toronto, and Vector Institute released Gamma-World, a multi-agent world model using simplex-based positional encoding and hub tokens to cut interaction cost from quadratic to linear, with 8-player latency dropping from 17.6 ms to 4.5 ms.

Why it matters: NVIDIA and Tsinghua researchers propose the multi-agent world model γ-World, using simplex positional encoding and hub tokens to cut inter-agent interaction cost from quadratic to linear.

The Verge · AI

Tech companies desperately want to film you doing chores

Shift said it would clean New Yorkers’ homes for free if it can film cleaners doing chores such as washing dishes, wiping counters, dusting tables, and mopping floors, creating domestic robot training data; the snippet does not disclose consent terms, data retention, pricing, or expansion timing for cities such as London.

Why it matters: Shift trades free house cleaning for video of New York households washing dishes and mopping floors, footage meant to train home robots, with plans to expand to London and other cities.

May 29Friday

The Verge · AI

This AI startup will clean your home for free to train future robots

Shift offers free home cleaning and records cleaners scrubbing, vacuuming, dusting, tidying, and washing to collect robot training footage; the RSS snippet does not disclose service cities, privacy terms, consent mechanics, or dataset scale.

Why it matters: Shift trades free house cleaning for footage of the work, turning ordinary chores into a robot-training data pipeline whose data value covers the service cost.

AI HOT (Curated Pool)

Tesla FSD Safety Claims Face Scrutiny

Tesla claimed FSD can be up to 10 times safer than humans, but Reuters found flaws in the comparison, with 11 traffic safety researchers saying Tesla used inappropriate baselines against broader federal crash data.

Why it matters: Reuters checked Tesla's safety report, and researchers plus former employees question the numbers behind its claims.

May 28Thursday

NVIDIA Blog

NVIDIA Research Advances Robotics From Simulation to the Real World

NVIDIA Research presented 8 ICRA papers on sim-to-real robotics: ScheduleStream delivered a 3x speedup for multi-arm planning, COMPASS reached about 80% success across 20 real-world navigation trials, and Grasp-MPC achieved about 75% real-robot grasping success.

Why it matters: NVIDIA's COMPASS navigates real robots after simulation-only training, hitting roughly 80% success across 20 trials, 4.5x a pure imitation baseline.

AI HOT (Curated Pool)

Mistral AI launches physics AI model for industrial engineering

Mistral AI integrated the Emmi AI team and launched a physics AI foundation model for industrial engineering, with the post saying it can learn from geometry, boundary conditions, or measurement data and predict full physical fields on a single GPU in seconds.

Why it matters: Mistral's physical AI foundation model for industrial engineering predicts full fields like temperature and stress in seconds on a single GPU, a route worth watching for industrial simulation.

Synced · WeChat

ICML 2026: AutoMoT reaches SOTA on Bench2Drive and nuScenes

NTU AutoMan Lab, Harvard, and Xiaomi Auto proposed AutoMoT, a unified VLA driving model using a 4B Qwen3-VL Understanding Expert and a 1.6B Action Expert with asynchronous inference, reaching 89.42 DS and 74.09% SR on Bench2Drive with AutoMoT+.

Why it matters: AutoMoT splits a 4B vision-language model and a 1.6B action expert into asynchronous inference, taking first place on both Bench2Drive and nuScenes, a split-model idea for end-to-end driving.

Synced · WeChat

Chinese pretrained embodied model Wall-OSS-0.5 is open sourced

X Square Robot open sourced Wall-OSS-0.5, a VLA model whose 400k pretraining checkpoint scored above 80 on 4 of 17 real-robot zero-shot tasks, with weights, code, training recipe, ablations, and a DMuon optimizer implementation released.

Why it matters: X Square Robot open-sources Wall-OSS-0.5, a vision-language-action model with a 400k-step pretrained checkpoint, weights, code and training recipe; 4 of 17 real-robot zero-shot tasks top 80% success.

May 27Wednesday

Synced · WeChat

From Foundation Models to Physical AI, Samsung Moves Into the Core LLM Race

Samsung disclosed three AI efforts—Meki, M2RL, and LiveClawBench—covering a memory-based edge architecture, multi-domain reinforcement learning, and Physical AI evaluation; the article also says Samsung has purchased tens of thousands of GPUs for AI infrastructure, but does not provide model size, training budget, or deployment timelines.

Why it matters: Samsung published three studies, Meki, M2RL and LiveClawBench, covering on-device memory, multi-domain reinforcement learning and physical AI evaluation, showing where it is placing bets on devices and robotics.

AI HOT (Curated Pool)

Reachy Mini enables fully local voice interaction

Reachy Mini implements local voice interaction through the speech-to-speech library, using a cascaded pipeline with a Realtime API-compatible WebSocket interface and default components including Silero VAD, Parakeet-TDT, and Qwen3-TTS.

Why it matters: Hugging Face built a fully local voice conversation pipeline for Reachy Mini, with each stage from Silero VAD to Parakeet-TDT to Qwen3-TTS documented, giving a complete path for anyone reproducing this on edge devices.

May 26Tuesday

New York Times Chinese

The Shared U.S.-China AI Anxiety: Being Harvested by the Future

Yi-Ling Liu compares U.S. and Chinese AI anxiety through labor, companionship, and agency: over 70% of U.S. teenagers report using chatbots as companions, while China is projected to reach 200 million single-person households by 2030.

Why it matters: The author sees the same unease about AI among ordinary people in China and the US, and shows why the gains don't reach individuals.

May 25Monday

Synced · WeChat

A 1B-Gaussian 3D World Runs in the Browser, Outperforming Fei-Fei Li’s Spark

Manycore Tech open-sourced Aholo Viewer, a browser-based 3D Gaussian Splatting viewer that used half the memory of Spark 2.0 in a 300M-Gaussian test, loaded 2x faster, rendered 3x faster, and supports scenes with up to 1B Gaussian points.

Why it matters: A browser-side 3D Gaussian splatting viewer with concrete memory, load and render numbers against Spark 2.0, plus a 1-billion-point claim.

May 22Friday

AI HOT (Curated Pool)

NDRC to Accelerate Embodied AI Training Infrastructure

China’s NDRC said humanoid robot race teams increased from more than 20 to over 100, with finishers rising from 6 to more than 40, and it will build embodied AI training infrastructure and pilot application bases for factories, malls, and homes.

Why it matters: The NDRC called for faster embodied-intelligence training infrastructure, citing Yizhuang half-marathon entries rising from 20-plus teams to over 100 and finishers from 6 to over 40, pointing robots from track to factory, mall and home.

Synced · WeChat

CVPR 2026 | HiF-VLA: A Motion-Centric World Action Model

Westlake University and collaborators introduced HiF-VLA, a motion-centric VLA framework that extracts compact Motion vectors with codecs such as H.264 and uses a joint expert to predict future visual motion and generate action sequences, reporting 31.4GB peak memory and 117.7ms latency under the cited history-window setting.

Why it matters: A Westlake University team proposes HiF-VLA, replacing per-frame visual encoding with H.264 motion vectors and predicting visual dynamics and actions together, at 31.4GB peak memory and 117.7ms latency.

Bloomberg Technology

Waymo Takes Multi-City Pause on Floods, Halts Freeway Access

Waymo temporarily halted robotaxi service in five cities because its vehicles may attempt to drive on flooded roads; the RSS snippet says the same issue recently triggered a recall of thousands of vehicles, but the post does not disclose the city list or restart timing.

Why it matters: Waymo paused service in five cities after cars drove into flooded Atlanta roads, the same recognition flaw behind a recall of thousands of vehicles.

NVIDIA Blog

NVIDIA GTC Taipei at COMPUTEX: Live Updates on What’s Next in AI

NVIDIA won four COMPUTEX 2026 Best Choice Awards for Vera Rubin NVL72, Jetson Thor, and Alpamayo; Vera Rubin NVL72 connects 36 Vera CPUs and 72 Rubin GPUs, and NVIDIA says it delivers up to 10x higher inference performance per watt and 10x lower cost per token.

Why it matters: NVIDIA took four Best Choice awards at COMPUTEX 2026 with Vera Rubin NVL72, Jetson Thor and Alpamayo, and NVL72's stated inference efficiency is 10x higher than the previous generation.

May 21Thursday

AI HOT (Curated Pool)

FSD officially launches in mainland China

The title says FSD has launched in mainland China, while the post only states an official entry into the mainland and does not disclose eligible cities, vehicle models, pricing, or regulatory conditions.

Why it matters: Tesla officially announced FSD's entry into mainland China, but gave no available cities, supported models or pricing, so the value here is that a long-awaited autonomous driving launch finally has an official line.

May 20Wednesday

Xinzhiyuan · WeChat

Behind Jensen Huang’s Douzhi Moment, Chinese GPUs Are Filling CUDA’s Moat

Moore Threads presented progress on its MUSA GPU ecosystem, with SDK 5.1.0 targeting CUDA 12.8 and supporting 761 driver and runtime APIs. The post says MUSA has entered SGLang’s mainline, is listed for 2026 Q2 hardware support, and supports automated library migration via MUSACODE.

Why it matters: Moore Threads SDK 5.1.0 targets CUDA 12.8 compatibility across 761 driver and runtime interfaces, SGLang mainline merged MUSA support, and hardware is slated for Q2 2026.

Xinzhiyuan · WeChat

UISEE Lists in Hong Kong as a Full-Scenario L4 Autonomous Driving Stock

UISEE listed on the Hong Kong Stock Exchange at HK$60.30 per share, with its public offering oversubscribed 6,777.29 times and a 90.5% share of the Greater China airport L4 commercial vehicle market in 2025.

Why it matters: UISEE listed on the Hong Kong exchange with 90.5% share of airport L4 commercial vehicles and 6,777x oversubscription, showing autonomous driving's progress in closed scenarios.

New York Times Chinese

China Wants to Expand AI, but Not at the Cost of Jobs

Chinese courts have publicly backed workers in AI-replacement dismissal cases for a third time; in the Hangzhou case, an employee’s monthly salary was cut from 25,000 yuan to 15,000 yuan before dismissal, and the court found the employer failed to provide a reasonable placement and had illegally terminated the worker.

Why it matters: A Hangzhou court ruled it illegal for a company to cut pay and fire a worker after replacing his QA job with AI, the third such public win for workers in China.

TechCrunch · AI

Google’s Genie world model can now simulate real streets with Street View

Google DeepMind is integrating Street View with Project Genie for interactive street-level simulations in robotics, gaming, and travel; the post does not disclose model parameters, launch timing, or evaluation results.

Why it matters: Wiring Street View into Project Genie gives robotics and game teams interactive 3D scenes built from real streets.

May 19Tuesday

AI HOT (Curated Pool)

Hyundai Motor Group Plans to Deploy 25,000 Boston Dynamics Atlas Humanoid Robots

The title says Hyundai Motor Group plans to deploy 25,000 Boston Dynamics Atlas humanoid robots; the post does not disclose the rollout schedule, deployment sites, or purchasing terms.

Why it matters: Hyundai Motor Group plans to deploy over 25,000 Boston Dynamics Atlas robots across Hyundai and Kia plants, targeting 30,000 units a year by 2028, a signal of humanoid scale in auto manufacturing.

AI HOT (Curated Pool)

Horizon Open-Sources 400M-Parameter Robot Control Model HoloMotion-1

Horizon Robotics Lab open-sourced HoloMotion-1, a 400M-parameter full-body humanoid control model that uses MoE sparse activation and KV-cache inference to reach about 300 FPS on-device, with code and a technical report released.

Why it matters: Horizon Robotics' lab open-sourced HoloMotion-1, a 400M-parameter whole-body control model for humanoid robots, reaching about 300FPS on-device via MoE sparse activation and KV-cache, well above the usual 50Hz control rate.

QbitAI · WeChat

Chinese GPU vendor Moore Threads releases MT Lambda for embodied AI simulation

Moore Threads released MT Lambda, an embodied AI simulation platform that combines physics, rendering, and AI engines, and demonstrated the robot dog “Xiaofei” executing a Sim-to-Real policy trained 100% in simulation on domestic hardware.

Why it matters: Moore Threads released the MT Lambda simulation platform, which combines physics, rendering and AI training on its own GPUs; the robot dog Xiaofei learned to walk in simulation and moved straight to hardware.

AI HOT (Curated Pool)

NVIDIA fine-tunes Cosmos Predict 2.5 with LoRA/DoRA for robot video generation

NVIDIA published a Hugging Face post on fine-tuning Cosmos Predict 2.5 with LoRA and DoRA to generate robot first-person videos from text prompts; the post does not disclose dataset size, training cost, or evaluation results.

Why it matters: NVIDIA's technical blog shows lightweight LoRA and DoRA fine-tuning of Cosmos Predict 2.5 to generate first-person robot footage like arm grasps from text prompts, a cheap path to training video for embodied AI.

May 18Monday

AI HOT (Curated Pool)

Baidu Core AI Business Revenue Exceeded RMB 13.6B in Q1

Baidu reported that its core AI-driven business generated more than RMB 13.6 billion in Q1 2026, up 49% year over year, and accounted for more than half of Baidu’s general business revenue for the first time.

Why it matters: Baidu's core AI-driven business passed 13.6 billion yuan in Q1 revenue, up 49% year on year and more than half of general business revenue, led by AI cloud, AI apps and Apollo Go.