July 28, 2026
Optimus? More like Optional
Transformer Transformer: A Unified Model for Motion-Conditioned Robot Co-Design
AI can now design a robot for one job, and commenters are already ditching humanoids
TLDR: Researchers built an AI system that can design a custom robot for a specific task and showed it working better than the original machine. In the comments, people immediately turned it into a fight over whether humanoid robots are overrated — with others joking that this is the first step toward robots designing their own kids.
A new robot paper just dropped, and the real action is in the reactions. The researchers say they built a system that can watch a sample motion, like flinging cloth, and then spit out a custom robot body built for that exact job. Not just the arm shape, but the whole package: joints, motors, and how it should move. In one real-world test, their made-to-order design did the task much better than the original machine, cutting motion error by 73% and easing joint strain too. In plain English: instead of forcing one robot to do everything badly, they want to build the right robot for the right gig.
That instantly lit up the comments with a spicy identity crisis for humanoid robots. One of the strongest takes came from esjeon, who basically declared this could "un-hype humanoids". Translation: if custom robots become cheap and easy, why keep pretending a human-shaped robot is the answer to every problem? That’s the big drama here — general-purpose robot vs. weird little specialist machine. And yes, the specialist crowd came in hot.
Then things got gloriously sci-fi. pk-protect-ai called it the start of "robotic auto-evolution", imagining self-improving machines designing their own upgraded bodies in a tiny factory. So the vibe swung from serious engineering breakthrough to "congrats, you invented robot reproduction" in record time. Even the paper’s punny name, Transformer Transformer, got a side-eye laugh. The mood was equal parts impressed, amused, and lightly terrified — which, honestly, is peak internet.
Key Points
- •The article presents Transformer Transformer, a unified model that generates complete robot embodiments from manipulation demonstrations for motion-conditioned robot co-design.
- •A fabricated design for cloth flinging on an ALOHA2 bimanual platform reportedly reduced tracking error by 73% and maximum joint speed by 30% versus the original robot.
- •The system uses a diffusion transformer trained on RoboTokens, a unified tokenization covering robot embodiment structure, state, and action data.
- •The same architecture is described as supporting multiple embodiment spaces and use cases, including embodiment generation and cross-embodiment control.
- •The method uses reward-agnostic dynamics predictions that are converted into reward-specific value predictions at inference time through Dynamics Self-Guidance, and experiments are said to improve performance and runtime over an evolutionary baseline.