Minds in Machines
Weekly Roundup #1
Fellow Constructor,
Welcome to the inaugural edition of the Minds in Machines Weekly Roundup. Its fantastic to have you here. In this letter I explore where the physical world and artificial intelligence intersect. Minds in Machines. You can expect stories, news, research and random musings. Any feedback is appreciated.
In the Spotlight:
Brain update.
Image credit: (NVIDIA)
NVIDIA has launched its $3,499 developer kit Jetson Thor, powered by Infineon microcontrollors, sensors and actuator technology. It delivers 7.5x compute with a 3.5x the energy efficiency improvement over Orin. In raw numbers Jetson Thor delivers up to 2070 FP4 TFLOPS of AI compute and 128 GB of memory.
Early adopters include industry leaders Agility Robotics, Amazon Robotics, Boston Dynamics, Caterpillar, Figure, Hexagon, Medtronic and Meta, while 1X, John Deere, OpenAI and Physical Intelligence are evaluating Jetson Thor.
Out in the wild Galbot has already integrated Jetson Thor into its G1 humanoid robot. Professor Wang He, Founder & CTO stated: “Our G1 Premium, now running on NVIDIA Jetson Thor, delivers significant gains in speed and real-time reasoning, enabling our proprietary VLA models to achieve enhanced real-world performance”. Galbot recently won a Gold Medal at the World Humanoid Robot Games and has deployed its robors in 10 pharmacies.
You can get your hands on Jetson Thor at a hackathon organised by Seeed Studio and HuggingFace. Its going to be exciting to see what is about to be built with this new mind.
Research that Caught my Attention:
Researchers at Tsinghua University, together with industry partners have published a paper on Memory in Vision-Language-Action Models for Robotic Manipulation: MemoryVLA.
The core problem: Robotic manipulation is non-markovian and therefore temporal context matters. Current approaches ignore this and this causes issues. The image frame before, during and after pressing a button looks remarkably similar.
The core idea: Add a dual memory system to VLAs: short-term “working” memory tokens + a long-term Perceptual-Cognitive Memory Bank then condition a diffusion action head on what the robot retrieves.
The results: Dual-memory VLA hits 84% across 12 real-world tasks (+26 pp long horizon vs SOTA) and 71.9/72.7/96.5% on SimplerEnv-Bridge/Fractal/LIBERO-5 (including +14.6 pp on Bridge)
All credit: MemoryVLA
Random Musing of the Week:
I listened to an older podcast with Eben Upton, Founder of Raspberry Pi this week. Couple of things got me thinking:
Features get added to the board through the following formula: (Cost of feature / Percent of people finding it useful). This means a $1 feature everyone finds useful gets same priority or 10c feature 10% of users find useful get same priority. This pushes lesser used features onto external hardware.
The Raspberry Pi Zero is his favourite Pi because it explores a different application of Moores Law. Most companies (including RPI) apply Moores Law in the following way: Pick a price point and then fit ever more compute every 12-18 months. The Pi Zero picks an amount of compute (2000era PC) and sell it as cheap as possible.
Incredible that this company was bootstrapped with just £100,000 investment.
Funding News:
Thematically, defence applications continues to draw funding both non-dilutive and venture.
Unmanned Ship Startup Blue Water Autonomy raises $50 million Series A (Fortune)
Reliable Robotics wins $17.4 million U.S. Air Force funding for autonomous aircrafts (Reliable Robotics)
Reframe Systems raised $20 million to build affordable homes powered by robots achieving a 35% cost reduction (Businesswire)
Launches, Unleashed and spotted in the Wild:
Robomart launches delivery robot with innovative $3 flat charging structure. The company is looking to roll-out in Austin later this year.
Harvard scientists have demonstrated a smart, wearable robot helping patients with upper-limb disabilities.
Researchers at the University of Surrey have developed new drones based on the agility of birds of prey to navigate urban environments.
Thank you for reading. If you like this, please share with one person you think could get value from it. Means the world.
Keep constructing.
Dominic