🌐 REALTECH News, March 2025
Your monthly technology guide on software defined physical industries
Hi all, it feels like ages ago that I published the monthly news format. The world has changed at least twice since then.
New world order.
The world is changing so fast—AI is upending every industry, and Trump is upending every geopolitical alliance. DeepSeek released R1 on the US inauguration day, a shot across the bow for the US, which had conceded leadership in manufacturing but believed it was way ahead of China in frontier AI. Game done changed.
AI is being applied across every critical industry. Every nation-state is now focused on building leading technology capabilities, capacity, and sovereignty. There has never been a more impactful time to invest in this space.
Hopefully today’s newsletter can provide you with some respite and context from the noise. I will post a longer piece on EU + security mid-week, I’d originally intended to update you here but it deserves its own thread. Stay tuned.
In today’s edition:
Google’s AI co-scientist and a “datacenter of Einsteins”
Apple and Meta jump into humanoid robots
VLAMs go open source on Hugging Face
A slew of growth stage defence tech rounds
📣 PSAs
🧑🏼🔬 Research: Alongside Dealroom, we have compiled 40+ pages of in-depth analysis on trends in REALTECH. We analysed 15k companies, 11k investors over 10 years - read it here
Thanks to everyone who commented on the quick links and thoughts newsletter I put out a few weeks ago. The feedback was overwhelmingly positive, so expect more of the same.
✈️ March travels, I will be in Zurich (March 12th & 16th) HMU if you’re around!
Top Stories
Home robotics, humanoids, big tech and embodied AI
Humanoid robots continue to heat up, with Apple and Meta now in the mix. Apple is rumoured to be in proof-of-concept stage, with autonomous robots for its smart home ecosystem. Apple has, uncharacteristically, released a flurry of robotics research this last month, likely to attract talent. Papers include ELEGNT, EMOTION and ARMADA - which cover robot manipulation, scene generation and emotive expressions for humanoids. Interestingly, Apple released a video (above) of a non-anthropomorphic robot working alongside a human operator across a number of household tasks. They may decide to avoid the biomimicry of humanoids, thus avoiding the actuation and locomotion nightmares and the (un)canny valley of full humanoid designs.
Meanwhile, Meta has just spun up a robotics unit inside Reality Labs, hiring Mark Whitten (ex-GM Cruise CEO) to lead the group. Unlike Tesla’s Optimus, Meta’s focus seems to be AI-first, pushing open-source robotics research forward. Meta’s strategy continues to be open source leading AI research, in order to collapse the potential value capture of AI rivals. A classic ‘commoditise your complement’ strategy. With NVIDIA, Google (DeepMind), and Tesla in the humanoid race, Big Tech’s fight for AI embodiment is now in full swing.
Agentic science - how LLMs are pushing research forward
There’s been a flurry of papers and announcements in autonomous science, an emergent field that leverages LLMs to create “agentic” software scientists. Google Research just published ‘Towards an AI co-scientist’ paper (blog post here). The co-scientist, developed with Gemini 2.0, is a virtual multi-agent research collaborator capable of literature analysis, hypothesis generation, experiment design, and iterative improvement through self-play-based scientific debates.
Notably, the AI co-scientist has demonstrated its ability to generate verifiable research outputs, including a drug candidate for acute myeloid leukemia that was validated through lab experiments. The co-scientist even solved a complex problem that had taken microbiologists a decade to solve in a few hours.
Similarly, a recent paper ‘Transforming Science with Large Language Models…’, analysed 148,000 research papers and surveyed 5,000 scientists. It highlights how LLMs are accelerating the research cycle while raising new ethical challenges, including AI-generated misinformation and risks to scientific integrity.
Hugging Face’s CTO, Thomas Wolf wrote a long post on X related to this, about having a “country of Einsteins sitting in a data center”: he posits that real breakthroughs come from asking bold new questions—not just filling in gaps between existing knowledge. Right now, AI is great at interpolation, but the real test will be whether it can challenge assumptions and generate entirely new insights
The US defence sector is getting a major shake-up
New US Secretary of Defense, Pete Hegseth is overhauling how the Pentagon buys and deploys tech, cutting through slow bureaucratic processes and shifting focus to AI, autonomous systems, and software-defined warfare. The Defense Innovation Unit (DIU) is set to take on a bigger role, accelerating procurement from startups and SMEs, while the traditional defense primes—long protected by inertia and connections—find themselves in the crosshairs. Alongside this, the DOGE initiative is pushing for greater transparency, signaling a move away from bloated legacy programs and primes to rapid prototyping and real-world testing:
“There’s a lot of programs around here that we’ve spent a lot of money on that, when you actually wargame it, don’t have the impact you want them to.”
At the same time, the US is reallocating resources from Europe to the Indo-Pacific, where its threat lies. A proposed $150 billion defense budget increase will fund naval expansion, integrated missile defense, and next-gen deterrence systems. AI-driven command-and-control, autonomous combat vehicles, and cyber resilience are at the top of the priority list. The message is clear: the Pentagon is abandoning its old, slow-moving model in favour of speed and technology—to the dismay of primes.
🤓 Stories you need to know
Shadow Robot’s DEX-EE robotic hand (Robot Report): their latest dexterous robotic hand was developed alongside DeepMind and features a tendon system and co-located sensors for improved learning
What Google’s return to defense AI means (DefenseOne): Google reversed its 2018 decision to stop selling software and AI to the Pentagon.
The Smart Machines Strategy 2035, an independent paper from the UK’s Robotics Growth Partnership, outlines a roadmap to position the UK as a global leader in robotics
TSMC unveils “$100bn” US chip fab investment (Financial Times): whilst TSMC may have rounded their headline number up, regardless, they continue to build leading edge fabs in the US. A key piece for the US to ensure AI sovereignty
Using open-source VLAM π on Hugging Face (Hugging Face): Physical Intelligence released their π vision-language-action model for general use. Hugging Face ported it to their LeRobot repo
Saronic plans autonomous shipyards in the US (Defense News). The company, which just raised a $600m round (more below), is building a fully autonomous shipyard. The US’s shipbuilding capacity has atrophied massively, and China’s shipbuilding ability is some 230 times greater than the US! Yikes.
EgoMimic, using Meta’s smart glasses for data collection (Meta): Georgia Tech has used Meta’s Project Aria glasses to collect ego-centric visual data for robot training
🔬Research
Robust Autonomy Emerges from Self-Play (Apple): GIGAFLOW, a large-scale self-play reinforcement learning framework, enables AI to master autonomous driving without human data, achieving state-of-the-art performance across driving benchmarks —averaging 17.5 years of continuous driving between incidents in sim
Intuitive Physics Understanding from Self-Supervised Video Learning (FAIR at Meta, Gustave Eiffel, EHESS): V-JEPA, a self-supervised AI model, learns intuitive physics by predicting missing parts of videos, outperforming traditional video and language models in detecting when physical laws are broken
Physical Intelligence launch ‘Hi Robot’ (Physical Intelligence): the system uses two-level inference, dubbed ‘System 1’ and ‘System 2’, similar to Figure AI’s Helix model and a nod to Daniel Kahnemann’s work, allowing the robot to better complete and ‘think’ through complex tasks
Improving Vision-Language-Action Model with Online Reinforcement Learning (Tsinghua University, Berkeley): iRe-VLA is an iterative reinforcement learning (RL) framework that enhances VLA models by alternating between supervised learning and online RL, to improve robotic decision-making
💰Notable Funding Rounds
Saronic ($600m Series C) the autonomous surface vessel company, raised a round led by Elad Gil
Apptronik ($350m Series A) the humanoid robot maker, raised from B Capital and Capital Factory, with participation from Google
Shield AI ($240m Series F-1) the autonomous air defence company, raised from L3 Harris and existing investors a16z
Verkada ($200 million Series E), the cloud-based physical security technology provider, raised funds led by General Catalyst
UVeye ($191 million funding), the provider of AI-powered vehicle inspection systems raised from Woven Capital
Augury ($73 million funding), the AI-driven machine diagnostics company, raised from Lightrock
Nomagic ($44 million Series B), the warehouse automation company, raised funds led by the European Bank for Reconstruction and Development (EBRD)
Viam ($30 million Series C), the robotics data platform, secured investment from Union Square Ventures and Neurone
Earth AI ($20 million Series B), the AI-driven mineral exploration company, raised from Tamarack Global and existing investor Cantos
Comand AI (€8.5 million seed) the French AI defense company, raised from Eurazeo
The link under "Thanks to everyone who commented on the quick links and thoughts " was a private page?: https://realtechnews.substack.com/publish/posts/detail/157386521?referrer=%2Fpublish%2Fposts%2Fpublished