Figure's $3.5B compute bet, Unitree's boxing claim
Figure stacked a $3.5B Nscale compute deal on top of Index, a network collecting task video from 100+ countries, and Unitree says its humanoids are now fighting each other with no human in control.
The big stuff
Figure stacks $3.5B Nscale compute deal on Index, its crowd-sourced task video network
Per MacNeil's breakdown, Figure announced two things a week apart. Index is a creator network pulling in real-world task video at roughly half an hour of uploads every second, from homes and workplaces in 100+ countries. Nscale is compute: $3.5B to start, intent past $6B, up to 100,000 NVIDIA Vera Rubin GPUs. Adcock has said Figure is largely constrained by data and compute, so MacNeil reads Index as the data bet and Nscale as the compute bet.
And then there are chapters money can buy. Figure has entered the chapter of scaling intelligence through data and compute
But finding the 1% of data that actually teaches the model is what compounds. Search and curation, not just capture.
Demos and launches
Unitree claims first fully autonomous humanoid robot combat, driven by its world model
Unitree says UnifoLM-X2-1.0, a world-action foundation model (a model that predicts what happens next and turns that into robot actions), now drives humanoids through fully autonomous combat with real-time planning. The company calls it a breakthrough in instant planning, decision-making, and dynamic interactive execution. It goes further and says the result validates the feasibility of large-scale deployment of world model-driven humanoids.
This validates the fundamental feasibility of large-scale deployment of world model-driven humanoid robots.
My @EvolveMMA membership will start to pay off.
Research
OopsieData passes 15k real robot episodes from 54 labs, failures included
Paul Zhou posted a progress update on OopsieData, a multi-lab effort to pool real-world manipulation rollouts (recordings of a robot attempting a task), including suboptimal successes and failures, across different robots and labs. As of last week 54 labs had signed up with 10k+ episodes; this week it stands at 15k+ episodes, about 140 hours. It started with 17 contributing labs and is still open for signups.
As of last week we had 54 signed up labs with 10k+ contributed episodes, and as of this week we have 15k+ episodes (140h).
Quick hits
Autowash's two-arm robot car wash goes commercial in Wheat Ridge
Lukas Ziegler reports that Autowash Robotics went into commercial operation this month at its Lakeside site in Wheat Ridge after about 14 months of development. The system scans each car in 3D, spoilers and cargo boxes included, then two industrial arms plan a touch-free cleaning path for that specific vehicle. Ziegler's numbers: under 20 gallons of water per wash, clean and dry in about four minutes, and a four-bay site drawing 320 amps, around half a conventional touch-free installation. He says he is curious how it holds up at scale.
Tesla's Ashok Elluswamy says Robotaxis work equally well in the rain
Elluswamy made the claim while quoting a rider's post about a Robotaxi trip through heavy rain in Tampa. That is the whole evidence base: one rider's account and a one-line company claim, with no data on rides, interventions, or failures in wet conditions.
RealSense launches Perception Studio for robot developers, with close-range vision
Lukas Ziegler, speaking with @chrismatthieu at Automate in Chicago, reports that RealSense launched Perception Studio alongside the D585 Pro camera. Three features to start: visual inertial odometry (tracking the robot's own motion from camera and IMU data, no wheel encoders needed), close-range vision down to 12 centimeters on a camera that normally sees 6 meters, and built-in people detection. RealSense is putting the tools out early to get developer feedback before the tech hardens.
Stanford's Oussama Khatib robotics lectures are free online
Lukas Ziegler flagged a free 16-lecture Stanford course from Oussama Khatib covering spatial transformations, forward and inverse kinematics, Jacobians, trajectory generation, motion planning, Newton-Euler and Lagrangian dynamics, manipulation and compliance, and whole-body control. If you want the math behind how robots move, Ziegler calls it a great starting point.