📡 News Roundup
1. Boston Dynamics Atlas Uses RL to Learn Heavy Lifting — Whole-Body Control at a New Level
Boston Dynamics has released the latest update on its electric Atlas humanoid: through reinforcement learning (RL) practiced in simulation, Atlas learned to lift and carry a mini-fridge using its entire body — arms, torso, and legs — rather than relying on fingertip manipulation alone. The team notes that this milestone was reached just weeks after Atlas’s public debut in January 2026, marking a leap in “whole-body physical intelligence” for humanoid robots.
đź”— Boston Dynamics: Training a Humanoid Robot for Hard Work
2. Hello Robot Unveils Stretch 4 — Practical Home Robot Without the Humanoid Hype
Hello Robot has officially launched Stretch 4, a pragmatic wheeled home robot that deliberately avoids the humanoid form factor. Key upgrades include an omnidirectional mobility base, a redesigned sensor head with dual hemispherical LiDAR and Luxonis stereo cameras, and an Intel compute platform. Co-founder and CEO Aaron Edsinger says Stretch 4 is designed to transition from a research platform to real home deployment — a pointed contrast to the current humanoid robot frenzy.
🔗 IEEE Spectrum: Hello Robot’s Wheeled Home Robot Ditches Humanoid Hype
3. ISO 13482 Gets Its First Major Overhaul in 12 Years — Safety Standards for Home Robots
IEEE Spectrum reports that ISO 13482, the safety standard for personal care robots, is undergoing its first major revision since 2014. Researcher Jae-Seong Lee of Korea’s ETRI points out that the current framework addresses basic hazards like collision detection but fails to account for the “bidirectional coupling” of human-robot interaction — how a robot changes human behavior and how humans alter what the robot perceives. As humanoids move from labs into real homes, this governance gap needs urgent attention.
đź”— IEEE Spectrum: Domestic Humanoid Robot Safety Standards Are Shifting
4. ICRA 2026 Kicks Off in Vienna — Global Robotics Community Gathers
The International Conference on Robotics and Automation (ICRA 2026) is underway in Vienna, Austria, running June 1–5. As one of the top academic conferences in robotics, this year’s agenda highlights humanoids, embodied AI, and reinforcement learning. Notable projects include the WiXus wheel-legged robot from the University of Tokyo’s JSK Lab, among many others.
đź”— ICRA 2026
5. MiniMax M3 Debuts — Chinese LLM Surpasses GPT-5.5 at 5–10% of the Cost
Chinese AI startup MiniMax has released its M3 large language model, which outperforms GPT-5.5 and Gemini 3.1 Pro on several key benchmarks at just 5–10% of the cost. M3 features a million-token context window, native multimodality, and planned open-weight release. For the embodied AI space, models like M3 could significantly reduce the cost of underlying LLM inference for robot cognition.
đź”— VentureBeat: MiniMax M3 debuts
6. Microsoft Unveils Surface RTX Spark Dev Box — Local AI Inference Gets a Hardware Boost
At Microsoft Build 2026, the company introduced the Surface RTX Spark Dev Box, a compact desktop powered by NVIDIA’s Blackwell RTX Spark SoC with 128GB of unified memory. Capable of running AI models exceeding 120 billion parameters locally at up to 1 petaflop of AI compute, the device aims to free developers from per-token cloud inference costs. For robotics teams running edge AI deployments and local model testing, this pricing model shift is worth watching.
đź”— VentureBeat: Microsoft debuts Surface RTX Spark Dev Box
7. Open-Source Robot AI Platforms Are Lowering the Barrier to Entry
IEEE Spectrum published a feature tracing how open-source AI tools — from ROS to platforms from Nvidia, Hugging Face, and Alibaba — are democratizing robotics development. Nvidia’s director of product for robotics, Spencer Huang, notes: “To get into robotics, you no longer need a Ph.D.” The open-source movement is rapidly turning robotics from a specialized discipline into a platform anyone can build on.
đź”— IEEE Spectrum: Open-Source AI Makes It Easier to Build Smart Robots
🔍 Video Highlight
Atlas Moves a Fridge: How Reinforcement Learning Teaches Humanoids to Use Their Whole Body
This week’s recommended viewing is Boston Dynamics’ latest Atlas training video. The robot practiced thousands of lifting variations in simulation — different weights, shapes, and poses of the fridge — before acquiring the ability to brace, lean, and carry using its entire body surface (arms, shoulders, knees, torso). The key breakthrough is not about “seeing” the fridge, but about real-time adaptation through tactile and force feedback during interaction. For industrial deployment of humanoid robots, this kind of “whole-body physical intelligence” matters more than multimodal perception alone.
đź”— Boston Dynamics: Training a Humanoid Robot for Hard Work
Daily updates tracking smart hardware and robotics frontiers.