Robotics Weekly

CoRL's bitter lesson fight got personal

Keerthana Gopalakrishnan told roboticists to adapt, Chris Paxton posted a towel fold in reply, and Robocurve's numbers say Gemini is the odd one out.

Keerthana Gopalakrishnan says roboticists at CoRL are in denial. Her argument is that the bitter lesson, the idea that general methods plus scale beat hand-built domain tricks, applies to robot control too, and that people in the field keep assuming their corner is the exception.

"Robotic control, as we are learning, is not special and is adjacent to so many frontier modeling problems," she wrote, listing multimodal reasoning, CAD, computer use and coding. Her sharpest line is that a model good enough to make slides is already a multimodal VLA (a model that turns camera images and text into robot actions). She also points out the irony. VLA researchers spent years saying generalization is the real test of intelligence, and now that frontier models generalize better than specialized robotics models, the pushback is loud.

Keerthana Gopalakrishnan
@keerthanpg
X
Robotic control, as we are learning, is not special and is adjacent to so many frontier modeling problems
Nov 11, 2026 · View on X

Paxton answers with a towel

Chris Paxton replied with a video and a straight face, telling worried roboticists to visit the RoboColosseum setup at the BitRobot booth because it made him feel a lot better. The clip is captioned as GPT6 astra folding a towel. The joke lands because RoboColosseum is the open evaluation that runs real robot trials, so the counterargument is not a position, it is a scoreboard.

The third voice in the thread is data. Robocurve PBC posted that general AI capability predicts robot capability, with one glaring exception. Remove Gemini models from the set and the correlation coefficient jumps from 0.20 to 0.96. In other words, across everything else the link between smart model and capable robot is almost nonexistent, and Gemini is doing something the trend line does not explain.

That feeds a separate complaint from Prime (Shengqu) Cai, who said he still does not get how Gemini is not leading across domains like robotics, Blender workflows and spatial tooling given its native multimodal advantages. Gopalakrishnan's full reply to him was four words. "you're not paying attention."

What is actually settled

Nothing, yet. Nobody in these posts has shown a frontier model beating a robotics specialist on a shared physical benchmark, and Paxton's towel video is a punchline rather than an evaluation. What both camps do share is RoboColosseum style live trials on real hardware, which means this argument has an exit. Run the models, count the successes, publish the numbers.

Robocurve PBC
@robocurve
X
Removing Gemini raises the correlation coefficient from 0.20 to 0.96.
Oct 15, 2026 · View on X
Chris Paxton
@chris_j_paxton
X
Here's GPT6 astra folding a towel
Nov 11, 2026 · View on X

Sound like you were at the demo

One email a day. Under three minutes. The demos, the deployments and the part the press release left out.

Get the next one by email

Robotics every day from the people building it. The demos, the deployments and the arguments worth your time. Every claim links back to the engineer.