Michael Black vs GPT-6 Astra; Agility's $1.8M revenue
Michael Black says GPT-6 Astra made his ECCV paper obsolete before he could present it, and senior vision researchers are pausing work to measure the model. Also: Kat Scott reads Agility Robotics' SPAC numbers, $1.8M revenue against a $140M operating loss.
The big stuff
Michael Black says GPT-6 Astra made his ECCV paper obsolete before he presented it
Michael Black is presenting VIGA at ECCV 2026 this week, a method that turns a single image into a 3D Blender scene using an agentic approach (an AI that plans and calls tools in steps), and he says it has already been surpassed, first by people using Claude Code and now by GPT-6 Astra. His argument: any paper at the conference is likely two years out of date, and in AI two years means irrelevant. His proposal is that every paper should include a section like Related Work reporting how current large models perform on the task, and that reviewers should judge on insight rather than technical novelty. Meanwhile Ken Goldberg noticed a painting from his UC Berkeley Droid setup in a scene that Shengjie Wang says GPT-6 rebuilt in MuJoCo from a single video, sourcing the Droid and Franka assets itself.
If you haven’t tried using Astra (or whatever comes next) to solve your problem, then you have not done your homework. This omission should be seen as negatively as not having a previous work section.
I've paused part of my group's research to evaluate Astra head-to-head on existing benchmarks. This moment demands a rethink: where we keep re-building instead of measuring, and which long-assumed immovable North Star problems have already been shaken loose.
I think large models are going to make step changes to robotics too (not just "robot-native models").
Demos and launches
Bleu humanoid reportedly learned machine tending from six demos, ran four days at 96%
Lukas Ziegler reports that the Bleu Robotics team trained a humanoid on site the day before a trade show opened, from six human demonstrations. The task is six steps with a real machine in the loop: pick a part from a box, place it in a laser engraver, close it, press two buttons to start the cycle, open it, retrieve the finished part, at roughly one-minute cycles. Ziegler says it then ran for four days at VivaTech at more than 80 cycles a day with a 96% success rate. Bleu describes itself as an AI company born on the shop floor, building universal software for deploying robots into factories.
It then ran for four days at VivaTech at more than 80 cycles a day, at a 96% success rate.
We all know that these days most of humanoid demos are usually staged. Rehearsed task, or controlled conditions.
Deployments and money
Kat Scott says Agility Robotics SPAC shows $1.8M revenue, $140M operating loss
Kat Scott posted what she says are Agility Robotics' figures ahead of its SPAC (a merger with a listed shell company as a route to going public): $1.8M in revenue on a $140M operating loss. She called Agility objectively one of the better performing humanoid companies, which is what makes the SPAC interesting. Steve Crowe pointed to a $300M order from one customer that Agility needs to come through. Scott's reply: given their numbers, that is about 700 robots amortized over five years.
$1.8M in revenue on $140M operating loss. Humaoid robotics is going great!
Not great! Seems Agility needs the $300M order from that one customer to come sooner than later.
Given their numbers thats like 700 robots amortized over five years.
Quick hits
Gravis Robotics raised $200M from SoftBank at over $1B valuation, Ziegler says
Lukas Ziegler visited Gravis in Zurich, spoke with CEO Ryan Luke Johns, and says the company raised a $200M Series A from SoftBank at more than a $1B valuation. Gravis adds an autonomy layer to existing heavy equipment, mainly excavators, and trains digging in simulation of soil dynamics, since every bucket of dirt changes the terrain the machine has to understand.
Yilun Du argues robots should generalize by inference, not just more demos
Yilun Du posted a research perspective, Generalization by Construction, on combining learned world models with inference so a robot can work out tasks it was never trained on. His example is placing a mug on a crowded table: learn separate evaluators for support, clearance, and reachability, using energy-based models (models that score how good a candidate is rather than predicting it directly), then search over placements that satisfy all three.
Deepak Pathak shows S1 following four different video prompts in one kitchen
Deepak Pathak posted S1 following four different video prompt recipes in the same kitchen, including one he calls far out of distribution: putting a plate in a toaster. Chris Paxton called it an important experiment for showing the system can actually be taught different tasks, while saying it is early days and steerability is a prerequisite rather than the finish line.