🚨🤖 Robotics Weekly Digest (June 4, 2025)
From sign language translation to chopstick dexterity, this week’s updates capture the growing breadth of intelligence we’re baking into machines. Let's break it down 👇
Physical Intelligence unveils “knowledge insulation” to speed up model training 7.5x. By integrating diffusion-generated outputs from VLMs during action model training, Physical AI improves both sample efficiency and language-following behavior. This technique allows for faster convergence in vision-language-action models, while preserving instruction integrity and grounding. Physical Intelligence
Figure’s F.03 humanoid takes its first steps, and its AI stack gets a revamp. Figure announced that its next-gen humanoid, F.03, is officially walking. Alongside this hardware progress, they revealed the formation of Helix, a newly merged AI division combining three teams under one umbrella.
DeepMind’s new Gemma models push into healthcare and accessibility. The latest drop from DeepMind includes SignGemma, designed to convert sign language into spoken words (coming later this year), and MedGemma, which spans both a 4B-parameter multimodal model and a 27B text-only variant. MedGemma is tuned for medical image and text understanding, aiming to improve reasoning in clinical contexts. X (formerly Twitter)
China unveils hotel-service humanoid for autonomous guest support. In a new demo, the Zerith H1 humanoid robot handles real-world hotel tasks like room cleaning, shelf restocking, and guest assistance without human intervention. The system highlights how robotics is expanding beyond research labs and warehouses into service industry environments. Interesting Engineering
HopeJr: a $3K open-source humanoid with 66 actuated joints, now live on GitHub. The Robot Studio and Hugging Face dropped HopeJr, an ultra-affordable bipedal platform capable of walking, grasping, and manipulating objects. Built with 66 DOF, it brings full-body control to hobbyist and research-grade humanoids. The full bill of materials and source code are publicly available. GitHub
MicroFactory shows a $5K robotic assembly rig with dual arms and plug-and-play skills. Built by Igor Kulakov, MicroFactory is a compact boxed-frame setup with two robotic arms, powered by vision and capable of learning tasks via visual demonstrations. With modular grippers and low-cost components, it performs tasks like screwing, soldering, and repetitive assembly—offering a glimpse of desktop-scale industrial automation. X (formerly Twitter)
Robotera’s STAR1 learns to use chopsticks, pushing precision control further. STAR1 demonstrates a difficult two-handed task—grasping and using chopsticks—with clean, deliberate motion. As fine manipulation becomes a new performance benchmark, this milestone shows how quickly robot hands are catching up to human dexterity. Interesting Engineering
ree teases humanoid platform with 26 DOF. A new teaser from Unitree Robotics previews a humanoid with 26 join DOF, suggesting a step forward in full-body locomotion and manipulation. While specs remain under wraps, the release hints at continued competition in affordable humanoid hardware.
ETH Zurich trains a robot dog to play badminton using only onboard sensing. In a new demo, a quadruped learns to rally a shuttlecock using a single reinforcement learning policy coordinating 18 degrees of freedom. With a 12.06 m/s swing speed and sub-400ms reaction time, the robot successfully lands 10 consecutive returns—all without external perception. Science
Morgan Stanley: Optimus is the key to unlocking the “Elonomy”. In its latest
TSLA 0.00%↑ report, Morgan Stanley outlines a thesis for “Elonomy”—Tesla’s AI-native product ecosystem. Analysts forecast deep integration between Optimus and other Musk ventures: xAI models powering Optimus and Cybercab, SpaceX connectivity in Cybertruck, Neuralink + Optimus assistive prosthetics, Tesla employees collecting POV training data for robot nets, and more. X (formerly Twitter)
From Figure’s new walking humanoid and DeepMind’s medical multimodal models, this week showed just how fast robotics and AI are converging. We saw robots mastering fine manipulation, open-source platforms becoming more accessible, and new infrastructure shaping how machines learn and act in the world. Follow us @openmind_agi for more weekly digests as we track the rise of decentralized AI, agentic systems, and embodied intelligence—one digest at a time.