Good morning, robotics enthusiasts. Boston Dynamics and Toyota just showed Atlas running on one AI model that controls its whole body and learns new jobs from human demos.
There's no hand-written code for each new skill, just a person showing the robot what to do. If teaching a humanoid gets that simple, how fast could robots pick up real factory work?
In today's recap:
Atlas learns whole-body tasks from human demos
Nvidia's Jetson Thor brings server-level compute to robots
Air Force pays Reliable Robotics to automate a Cessna
BOSTON DYNAMICS
Atlas humanoid learns new tasks from human demos
Recaply: Boston Dynamics and Toyota Research Institute just demoed Atlas running on a Large Behavior Model, a single AI model that controls the whole humanoid. New skills come from human demonstrations, not hand-written code.
Key details:
Atlas's model turns head-camera images, body sensing and a language prompt into whole-body actions 30 times a second. It treats the hands and feet almost identically, instead of splitting walking from grasping.
Its policy is a 450M-parameter diffusion transformer that controls the full 50-degree-of-freedom Atlas. On an upper-body test stand, one model also tied rope, spread a tablecloth and moved a 22lb car tire.
Researchers threw in surprises mid-task, like closing a box lid and sliding the box across the floor, and Atlas adjusted on its own, Toyota said.
Boston Dynamics and TRI started the partnership in October 2024 and shared the video on August 20. Next, they plan to scale up data collection and test new algorithms.
Why it matters: Humanoids that did this before usually split walking and balance from arm control, which makes long, messy jobs hard to program. One model that learns from demos changes that: if a person can show it, Atlas can learn it. TRI's Russ Tedrake says the models need fewer demos as they get stronger. It's still research, though, and Atlas remains in development.
NVIDIA
Nvidia's Jetson Thor robot brain goes on sale
Recaply: Nvidia just launched Jetson AGX Thor, a computer module it calls a robot brain, with 7.5x more AI compute than Jetson Orin. It's built to run large AI models right on the robot.
Key details:
Built on Nvidia's Blackwell GPU, Thor runs vision language and vision language action models in real time on the robot. It handles live streams from many cameras and sensors at once.
Jetson Thor packs 128GB of memory, with 3.1x the CPU performance and 2x the memory of Orin. Its developer kit costs $3,499, and T5000 modules cost $2,999 each in orders of 1,000.
Robotics is still about 1% of Nvidia's revenue, CNBC reported. Its combined auto and robotics unit posted $567M in quarterly sales in May, up 72%.
Nvidia put Jetson Thor on general sale on August 25. Agility Robotics plans to use it in its sixth-generation Digit, and Boston Dynamics is building it into Atlas.
Why it matters: Humanoids need to see, reason and move at the same time, with very little delay. Thor puts server-level compute inside the robot, and Agility and Boston Dynamics are already signing on. Nvidia's Deepu Talla said the company doesn't build robots; it supplies the computers and software. Robotics is small for Nvidia today, but Jensen Huang calls it the biggest growth chance outside AI.
RELIABLE ROBOTICS
Air Force orders a self-flying Cessna cargo plane
Recaply: California startup Reliable Robotics just secured a $17.4M US Air Force contract to automate a Cessna 208B cargo plane. Its system handles taxi, takeoff, flight and landing with no pilot on board.
Key details:
Reliable's Autonomy System is a drop-in kit for existing, type-certified aircraft, and it isn't tied to one model. Layers of redundancy and advanced navigation aim for the reliability uncrewed flight needs.
Air Force money covers buying, installing and testing the system on a 208B freighter called the Super Cargomaster. Reliable, founded in 2017, has done Air Force autonomy work since 2021, including the KC-135 tanker.
Reliable's David O'Brien told AIN that "Pacific operations are ultimately anticipated," with Afwerx, an Air Force innovation branch, helping pick the first test site.
Work starts at the end of 2025, with R&D in the first half of 2026 and operations planned for 2026 to 2027, O'Brien said.
Why it matters: Getting supplies to scattered bases quickly is a core Air Force problem. Former Air Mobility Command chief Gen. Mike Minihan called autonomous aircraft "a true force multiplier" that cuts risk and widens where supplies can go. Reliable also plans to sell its kit commercially, and the FAA has accepted certification plans for an early version. Pilotless certification is still the end goal, not a done deal.
NEWS
Beyond the big stories in robotics today
Telexistence announced a Motion Data Factory that uses multi-joint robots to generate robot motion datasets for customers. Pre-orders open in September, the service starts in January 2026, and pricing depends on dataset length and task complexity.
FieldAI raised $405M across two rounds from backers including Bezos Expeditions, Nvidia's NVentures and Khosla Ventures, at a $2B valuation, CNBC reported. Its "physics-first" Field Foundation Models run quadrupeds, humanoids, wheeled robots and vehicles without maps or GPS.
Boston Dynamics' Spot robot dogs were eliminated in the quarterfinals of America's Got Talent Season 20. Five Spots danced to the Beach Boys' "Good Vibrations," a routine built with the company's Choreographer software.
Walmart will install Ranpak AutoFill systems across its five Next Generation Fulfillment Centers under a multi-year deal. Ranpak's AI vision measures the empty space in each box, then the machine adds just enough paper fill and seals it.
Rebl Industries signed framework deals with H&M and IKEA for its AI-powered warehouse robots, offered as a service where customers pay for work done. It also named ex-Volvo Cars executive Nicholas Tengelin as its first external CEO.
Harvard researchers built a soft wearable robot that learns each user's movements to help stroke and ALS patients eat and drink. Trained on a user's own data, it told shoulder movements apart with 94% accuracy.
Noetix Robotics won two golds and a silver at the World Humanoid Robot Games in Beijing, placing third on the medal table. Its E1 won the standing long jump at 1.25 meters, and its N2 scored 41.60 in floor exercise.
RealMan launched a 3,000-square-meter humanoid robotics data training center in Beijing, with 108 robots across ten real-world settings like eldercare and car assembly. It expects to produce over 1 million data points a year.
Hyundai Motor Group raised its planned US investment to $26B from 2025 to 2028, up $5B from March. It includes a robotics facility built for 30,000 units a year, and the whole plan should create about 25,000 jobs.
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason






