Walden Robotics, a spinout from Toyota Research Institute, launched in January 2026 and raised a $300 million seed round at a $1.1 billion valuation, co-led by Toyota and Deviation Capital with NVIDIA, Boeing, Samsung Ventures, and others participating. By February 2026 — under two months after spinning out — its robots were performing real production work at a Toyota plant in North America. The company is commercializing roughly a decade of TRI research, including Diffusion Policy and Large Behavior Models, and is building a full stack covering hardware, software, physical AI, and deployment. Its rapid transition from research spinout to live factory deployment signals a new pace for industrial robotics entering production environments.
1X Technologies has unveiled new tendon-driven robotic hands for its NEO humanoid robot, featuring 25 degrees of freedom, IP68 waterproofing, up to 45N finger force, 17.75Nm wrist torque, ±0.2mm position accuracy, and a rated reliability of over 2 million cycles. The hands are fully backdrivable and include multi-modal tactile sensing for pressure, location, and slip detection. 1X states the design is intended to match or surpass human hand dexterity so that AI training data — not hardware limitations — becomes the primary bottleneck for expanding robot capabilities. If the specs hold in real-world deployment, this represents a meaningful step toward humanoid robots performing the full range of manual tasks currently requiring human workers.
1X Technologies, the AI and humanoid robotics company behind NEO, has appointed Bill Nash as Chief Financial Officer. Nash brings financial leadership experience from companies in autonomy, aerospace, and life sciences. 1X is based in Palo Alto and develops humanoid robots designed to perform household chores and provide personalized assistance. The executive hire signals 1X is scaling its financial operations as it pushes toward commercializing its NEO humanoid platform.
Sakana AI, a Japanese AI lab, has launched Sakana Fugu, a multi-agent orchestration system that presents as a single OpenAI-compatible API endpoint while internally handling model selection, delegation, verification, and synthesis across a pool of expert agents. Two tiers are available: Fugu for balanced performance and low latency, and Fugu Ultra, which the company claims matches frontier models Fable and Mythos on engineering, scientific, and reasoning benchmarks. A key differentiator is its design around export-control resilience — swappable agents allow organizations to route around vendor restrictions, positioning it as a sovereign AI option for critical infrastructure. The launch signals a strategic shift from monolithic LLMs toward orchestration-layer models with geopolitical and operational advantages.
The US government issued an export control directive on June 12, 2026, ordering Anthropic to suspend all access to its Fable 5 and Mythos 5 models for any foreign national, forcing Anthropic to disable the models entirely for all customers to ensure compliance. The government's concern centers on a reported jailbreak technique, which Anthropic says amounts to asking the model to read a codebase and fix software flaws — a capability Anthropic argues is already widely available in models like OpenAI's GPT-5.5. Anthropic disputes the action, stating no universal jailbreak has been found and that their defense-in-depth strategy makes Fable 5's risks comparable to other deployed industry models. The order sets a significant precedent that could, by Anthropic's own assessment, halt all new frontier model deployments if applied uniformly across the industry.
Shift launched a service offering free apartment cleaning in New York City where vetted human operators clean homes while wearing recording devices to capture training data for robotics AI. Users pay nothing while the company monetizes the anonymized recordings of daily task performance to fund the service. The company plans to expand globally to handymen, repairs, and errands. This represents a novel business model that turns the gig economy into a data collection mechanism for embodied AI development.
Anthropic has signed an agreement with SpaceX to access Colossus 1, one of the world's largest AI supercomputers featuring over 220,000 NVIDIA GPUs and more than 300 megawatts of compute capacity. The partnership immediately doubles Claude Code rate limits for subscribers and removes peak-hour restrictions while raising API rate limits for Opus models. The deal also includes plans for future orbital AI compute capacity development, addressing the growing demand for AI training and inference that is outpacing terrestrial infrastructure capabilities.
SubQuadratic has launched SubQ, the first frontier LLM built on a fully sub-quadratic sparse-attention architecture that supports a 12 million token context window. The model is 52x faster than FlashAttention at 1M tokens and costs less than 5% of Anthropic's Opus (~$1.50/M vs $15/M tokens) while achieving competitive performance on benchmarks like SWE-Bench Verified (81.8%). SubQ uses content-dependent sparse attention that focuses only on relevant token relationships, reducing compute by nearly 1,000x compared to standard transformers. This breakthrough could fundamentally reshape LLM economics by making long-context processing practical and affordable for enterprise applications.
Meta Superintelligence Labs released Muse Spark, a natively multimodal AI model that combines visual and language reasoning with tool-use capabilities and multi-agent orchestration. The model is immediately available through Meta AI app and select partner APIs, with competitive performance claimed across multimodal perception, reasoning, and health applications including visual troubleshooting and medical explanations developed with 1,000+ physicians. This marks Meta's major re-entry into frontier AI competition after a nine-month infrastructure overhaul, positioning against OpenAI's GPT and Google's Gemini with plans to open-source future versions.
Willow released Atlas 1, a new speech-to-text model that achieves 1.2% word error rate on clean audio and 2.1% in real-world conditions, significantly outperforming competitors like OpenAI, Deepgram, and ElevenLabs who typically score 5-7% on clean audio and 10-15% in production. The model is built on what Willow claims is the first scalable human-powered transcription infrastructure for real-time dictation, developed over 14 months with $8.4M investment and 947 specialists. This breakthrough in accuracy could disrupt traditional typing by making hands-free dictation viable for professional productivity applications.
Google DeepMind launched Gemma 4, a family of open AI models including 31B Dense, 26B MoE, and mobile-optimized E4B/E2B variants that can run locally on personal hardware. The models feature native tool use for autonomous agents, up to 256K context length, multimodal capabilities, and are released under the fully permissive Apache 2.0 license. Performance benchmarks show dramatic improvements like Codeforces ELO jumping from 110 to 2150, with the 31B model ranking #3 on Arena leaderboards. This democratizes access to state-of-the-art AI capabilities for developers and enterprises without dependence on cloud APIs or restrictive licensing.
Asimov has launched OpenCrane, an open-source crane system designed specifically for humanoid robots. The frame stands 200 cm tall with a 180 cm span, constructed from 6063-T5 aluminum extrusions and features a modular design that allows mounting components without drilling. The complete specifications and assembly files are available on GitHub for free use by the robotics community. This addresses a key infrastructure need for DIY humanoid developers, providing a standardized frame that enables rapid prototyping and testing.
Chinese company Unipath has deployed a household robot that is currently in real-home use, capable of waking users on schedule, operating home appliances, organizing storage spaces, and cooking meals autonomously. The wheeled, non-humanoid robot with manipulator arms can perform tasks like cracking eggs, using spatulas, operating coffee makers and stoves, and cleaning. This represents a shift from prototype demonstrations to actual consumer deployment of multipurpose domestic robots that could reshape household labor and accelerate the global robotics competition.
President Trump has appointed 13 tech leaders to the President's Council of Advisors on Science and Technology (PCAST), including Meta CEO Mark Zuckerberg, Nvidia CEO Jensen Huang, Oracle Chairman Larry Ellison, Google co-founder Sergey Brin, and AMD CEO Lisa Su. The council, which could ultimately include 24 people, will advise on AI policy and other technology issues, with particular focus on competing with China's state-backed AI capabilities. This signals a direct industry-to-government pipeline for shaping national AI strategy and technology policy at the highest levels.
Figure AI's Figure 03 humanoid robot demonstrated advanced dexterity by catching packages thrown by human workers and sorting them by size, color, and shape into bins in a warehouse setting. The 56-second video shows the robot handling dynamic interactions including rapidly thrown packages while maintaining precise manipulation and balance. Salesforce CEO Marc Benioff's endorsement with over 1M views signals growing industry validation that humanoid robots are approaching viability for labor-intensive logistics jobs, potentially disrupting warehouse employment while advancing general-purpose robotics.
Amazon acquired New York-based Fauna Robotics, founded in 2024 by former Meta and Google engineers, bringing 50 employees and their $50,000 Sprout robot to the company. Sprout is a 3.5-foot tall, 50-pound bipedal humanoid powered by NVIDIA Jetson Orin that can handle voice interaction, object manipulation, and memory formation, with early customers including Disney and Boston Dynamics. The acquisition positions Amazon to compete directly with Tesla's Optimus and Figure AI in the consumer humanoid market. This deal could accelerate the path from expensive research platforms to affordable household robots by leveraging Amazon's supply chain and consumer reach.
Brett Adcock, founder of Figure AI and other ventures, has unveiled Hark Labs after 8 months in stealth, building what he calls the 'world's most advanced personal intelligence.' The system aims to create highly personalized AGI with persistent memory that can listen, talk, see, and proactively think like or ahead of users, paired with next-generation hardware as a universal human-machine interface. The lab is hiring across AI foundation models, infrastructure, computer use agents, and hardware engineering with major updates promised this year. This represents a significant entry into the personalized AGI space from a proven founder who has raised hundreds of millions for previous ventures.
Agile Robots SE and Google DeepMind have formed a strategic research partnership to integrate DeepMind's Gemini Robotics foundation models with Agile Robots' industrial robotics platform. Agile Robots has already deployed over 20,000 robotics solutions worldwide and will combine their German-developed hardware with DeepMind's AI models to create adaptable, reasoning robots for industrial environments. The partnership focuses initially on high-value industrial and manufacturing use cases where reliability and scale are critical. This collaboration aims to create autonomous, intelligent production systems that can transform entire industries through improved AI-powered automation.
Figure AI's Figure 03 humanoid robot demonstrated autonomous parcel sorting capabilities in a warehouse setting, picking up and organizing mixed packages from a conveyor belt while maintaining balance and dexterity even when faced with human interference. The 56-second demo video, shared by Salesforce CEO Marc Benioff, garnered nearly 1 million views and showcased the robot's precise hand movements and two-legged stability during complex manipulation tasks. This represents significant progress in general-purpose humanoid robotics for logistics applications, potentially disrupting labor-intensive roles in package handling and warehouse operations.
Anthropic launched computer use capabilities for Claude, allowing the AI to directly control macOS computers by opening applications, navigating browsers, and filling spreadsheets. The feature works through Claude Cowork and Claude Code, prioritizing connected app integrations like Slack and Calendar before requesting permission for direct screen control. Users can assign tasks remotely via mobile and set up recurring automation like daily email scans or weekly reports. This represents a major leap toward AI agents that can perform complex desktop workflows independently, potentially automating significant portions of knowledge work.