Michael Black says Astra made his ECCV paper obsolete
A senior vision researcher argues any paper at this year's conference is two years out of date, and Danfei Xu says robotics is next.
Michael Black is presenting VIGA at ECCV 2026 this week and says the paper is already out of date. VIGA takes a single image and outputs a 3D Blender scene, using an agentic approach (an AI that plans and calls tools in steps). It was the first method to do that, Black says. Then it got rejected once, publication slipped, and people using Claude Code passed it. Now, he says, GPT-6 Astra blows away all previous results.
The thread started small. Replying to Yuliang Xiu, Black wrote that he wondered why they are presenting VIGA at all given how quickly the field has moved. Xiu said he has paused part of his group's research to evaluate Astra head to head on existing benchmarks. Black said he would write more. Then he posted a long essay.
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.
The proposal
Black's arithmetic is blunt. A paper you read today rests on ideas at least a year old, which rested on literature that was already a year old, so anything at ECCV is likely two years stale. In AI, he writes, two years means your work is likely irrelevant.
His fix is a change to how papers are written. Every project should start by trying hard to solve the problem with existing tools, and every paper should carry a section analogous to Related Work where the related work is current large models and how they perform on the task. If the models fail, the solution should offer insight that outlives the next model release. He also wants reviewers to stop rewarding technical novelty, which he says pushes people to tweak architectures rather than move the field.
Robotics is not exempt
Danfei Xu replied that large models are going to make step changes to robotics too, not just robot-native models, and that researchers should either stick to hard problems and stay flexible on tools, or study the best tools and how they reshape the problem space.
The same day, Shengjie Wang posted that GPT-6 spun up a VM in the ChatGPT web UI, set up MuJoCo, sourced Droid and Franka assets including the meshes, and rebuilt a scene from a single video, built by Guanghe Li. Ken Goldberg replied that the painting in the background is from his Droid setup at UC Berkeley. Wang calls that proof that real2sim2real works. That is one post and one video, not a benchmark, and it follows other Astra claims circulating this week.
What makes this story different from a demo is who is talking. Named senior researchers are changing what they work on and what they want conferences to ask for.
I've paused part of my group's research to evaluate Astra head-to-head on existing benchmarks.
I think large models are going to make step changes to robotics too (not just "robot-native models").