Eidon AI shuts down: robot data has no buyers
Ted Xiao says the robotics data market has a single digit number of real buyers, and one company just closed because of it. Meanwhile Jiafei Duan posted unedited real-time footage of four frontier models running robot arms, and the field's verdict was: slow.
The big one
Eidon AI shuts down and the field agrees robot data has almost no buyers
Sam Padilla is shutting down Eidon AI after two-plus years collecting robotics data, saying the thesis was right but the business is brutally hard, and open-sourcing everything the company built. Ted Xiao says there is a brief arbitrage window but the market is saturated with a single digit number of real buyers.
the market is saturated, entry costs are surprisingly non-trivial, and there are a single digit number of real buyers
Shown this week
Duan runs Astra, Opus, Grok and MolmoAct2 in real time, no edits
Jiafei Duan followed his slow-Astra clip with a four-way real-time comparison of GPT-6 Astra, Opus 5.5, Grok 4.7 and MolmoAct2, zero-shot and with no cherry-picking. Chris Paxton's verdict on the Astra rollout: incredibly slow for a simple pick and place, and a sign that robotics benchmarks are still too easy. Duan says LLM robot control is progressing but low-level control still matters.
Incredibly slow for a very simple pick and place manipulation task that you could probably vibe code faster with SAM.
Black Forest Labs' 7B FLUX 3 Action tops RoboLab and ships open weights
Black Forest Labs released FLUX 3 Action, an open weights 7B world action model (predicts video and robot actions together) it says takes first place on the RoboLab benchmark, beating the previous best open model by 6.1 points with 56% fewer parameters and up to 3.95x faster.
Berkeley: one human demo, any multi-fingered hand, zero-shot sim-to-real
Tara Sadjadpour and Berkeley collaborators posted a method that turns a single human demonstration into a zero-shot sim-to-real visuomotor policy (trained in simulation, run on the real robot) for any multi-fingered hand. Jitendra Malik framed it as humans acting as humans plus robots training in simulation, and ended his post with "Whither tele-op?" His pitch: visual imitation then trial and error, the way children learn.
Teleop and other arguments
Adcock says data and compute gate humanoids, Goldberg asks if data is dead
Brett Adcock posted a four-stage model of humanoid maturity, where hardware and whole-body architecture cannot be brute-forced with capital, and scaling intelligence then manufacturing will take tens and eventually hundreds of billions of dollars. He says humanoids will need far more data and compute than LLMs. Ken Goldberg, at IROS 2026, asked the opposite question: could agentic robotics with pretrained LLMs replace the quest for demonstration data entirely?
Could agentic robotics with pretrained LLMs replace the quixotic quest for demonstration data?
Also on the timeline
Skild AI's humanoid plays football after 140 years of self-play
Skild AI says it trained its football robot through 140 years of self-play in a virtual World Cup. Chris Paxton called it dynamic, interactive whole body control with a human.
Paxton says mobile manipulation on Digit v4 worked first try
Chris Paxton ran end-to-end mobile manipulation on an Agility Robotics Digit v4 at IROS 2026. He says it worked out of the box, first time.
Frequently asked questions
Why did Eidon AI shut down?
Sam Padilla shut down Eidon AI after two-plus years of collecting robotics data, saying the thesis was right but the business is brutally hard. The company found that the market has almost no buyers for robot data, with only a single digit number of real buyers, because labs generate their own data instead.
What is a world action model?
A world action model is a type of AI system that predicts both video and robot actions together. Black Forest Labs released FLUX 3 Action, a 7 billion parameter open weights world action model.
Can robots learn from just one human demo?
Yes, according to Berkeley researchers, a single human demonstration can be turned into a zero-shot sim-to-real visuomotor policy (a policy trained in simulation and run on real robots) for any multi-fingered hand. The method combines visual imitation with trial and error, similar to how children learn.
What does sim-to-real mean?
Sim-to-real means training a robot policy in simulation and then running it on a real robot without additional changes. This approach can reduce the need for physical training data.
How much data and compute do humanoids need?
Brett Adcock posted that humanoids will need far more data and compute than large language models, and that scaling intelligence and manufacturing will eventually take hundreds of billions of dollars. However, Ken Goldberg asked whether agentic robotics with pretrained language models could replace the need for demonstration data entirely.