Current robot evaluations test software control policies but don't measure what existing hardware is actually capable of achieving. The author proposes dedicated hardware evaluations using any control method—autonomous models, teleoperation, or task-specific programming—to establish a lower bound on capabilities that improved AI models could unlock without hardware changes, revealing a potential "hardware-capability overhang" where existing robots can do more than current autonomous policies demonstrate.
Social media discussions debate AI capital expenditure sustainability, with concerns about return on investment, human dependency, and job creation. Commenters highlight China's hardware advantages, data scarcity as a bottleneck for embodied AI, and historical parallels to previous capex bubbles.
Boston Dynamics developed a new 13-degree-of-freedom hand for its Atlas humanoid robot, designed to manipulate objects with dexterity comparable to human hands while balancing strength, ruggedness, and manufacturability for industrial work. The four-fingered design with an opposable thumb uses direct actuation and tactile sensing to enable complex manipulation tasks including tool use, prioritizing function over replicating human anatomy.
Boston Dynamics has unveiled a redesigned hand for its Atlas robot optimized for mass production, trading humanlike aesthetics for reliability, manufacturability, and cost-effectiveness. The new hand features four fingers with 13 degrees of freedom, larger actuators, and the ability to use tools and perform diverse manipulation tasks. Unlike competing anthropomorphic designs, this pragmatic engineering approach prioritizes real-world deployment over cosmetic similarity to human hands.
Astra provides a production serving platform for robot foundation models, offering low-latency inference through either cloud or self-hosted deployment with predictable cost controls. The startup addresses the constraint that serving capable robot models reliably and affordably is more critical than further model training for practical robotics applications.
Runway announced Praxis-1, an open-weight world action model for robotics that leverages large-scale video pretraining to enable robot control across different embodiments. The model addresses the scarcity of real-world training data by learning from video, with early testing underway at partners like Noble Machines and Standard Bots before public release.
Gap is a robotic system for Variational Automation tasks where robots adapt to changing object geometries and poses within fixed workcells. It decomposes natural-language task descriptions into computation graphs using skill agents, then refines them through simulation rehearsal before deployment on real robots, achieving high success rates on tasks like grocery packing and popcorn making.
A programmer built a low-cost robot arm (SO-ARM101) for $353.30 using affordable servo motors and trained it with AI to pick up objects by demonstrating the task 50 times, leveraging Hugging Face's LeRobot library and the ACT policy model to enable intuitive robot control.
A developer equipped an OpenClaw AI agent with a physical robot arm from HuggingFace's LeRobot project, demonstrating that AI models can now configure robotic systems, perform object manipulation, and train other models with minimal manual intervention. The "code as policy" approach, which uses AI-powered coding to control robots, has advanced significantly and is gaining adoption across research labs, with Google DeepMind's Gemini outperforming other models on robotics benchmarks.
X posts discuss RWA tokenization infrastructure and Physical AI projects. Axis Robotics, a robot training data engine on Base chain, completed an oversubscribed presale after raising $12 million in July 2026. Posts highlight tokenization platforms like Brickken and stablecoin solutions as critical infrastructure for the RWA market.
Maxwell, an embodied AI model developed by China's Institute of Artificial Intelligence for Industries, achieved the highest score (91.9) on the Meta-World benchmark for robot physical tasks, surpassing competitors from Google DeepMind, Physical Intelligence, and other leading teams. The model can perform over 200 tasks without fine-tuning and demonstrates China's advances in embodied AI for robotics.
Stephen Fry delivers a lecture on artificial intelligence at King's College London, discussing AI as both promising and threatening. Drawing parallels to converging technological currents forming a tsunami, he argues for preparedness across universities and institutions, while humbly acknowledging the limits of human understanding regarding AI's future impact.
Rho is a foundation model for vision-language-action robots that separates adaptation into two stages: first learning a specific robot's embodiment, then adapting to particular tasks. This two-stage approach reduces the finetuning data needed by half compared to baseline models, with midtrained variants matching or outperforming competing systems on three physical robots.