Dyna unboxed Taku at CoRL and ran it zero-shot
Chris Paxton called it probably the most impressive demo at the conference.
Dyna Robotics says it took its Taku robot out of the box at CoRL and put it straight to work, doing long-horizon, whole-body mobile manipulation zero-shot, with no on-site data collection and no post-training.
The model driving it is Dyna-2.1. The robot is running live at Booth 500 in Griffin Hall, which is the part of this that matters most. A demo you can walk up to at a conference venue is a different kind of claim from a demo you watch in a cut video.
Why zero-shot is the claim
Unpack the jargon and the pitch is simple. Long-horizon means a task with many steps strung together rather than one grasp. Whole-body mobile manipulation means the robot drives around and uses its base and arms together instead of standing still in front of a table. Zero-shot means the policy was not given any data from this specific room, this specific lighting or this specific set of objects before it ran.
That last part is what robot teams usually quietly skip. Collecting a few hundred demonstrations on site, or fine-tuning overnight on the venue's furniture, is normal practice for a booth demo. Dyna says it did neither.
Who vouched for it
Chris Paxton, who was at the conference, called it "probably the most impressive demo at CoRL" and said long horizon mobile manipulation will always be the coolest thing you can do.
probably the most impressive demo at CoRL
Jason Ma replied to him with the line that frames why anyone should care. Long horizon mobile manipulation, he said, is what separates demos from real deployments. A robot that can pick one object in one pose is a party trick. A robot that can move through a space and chain steps without a human resetting it is the thing warehouses and back-of-house kitchens actually need.
Shown versus claimed
The zero-shot part is Dyna's claim. Nobody outside the company has audited what Dyna-2.1 was trained on, how many attempts the robot needed, or whether anything was taken out of the footage. Paxton saw a robot working at the venue and was impressed by it, which is a real data point, but it is not the same as verifying that no on-site tuning happened.
The open invitation is the strongest evidence here. Dyna is telling conference attendees to come stand in front of the robot for the length of the event. That is a hard setting to fake, and if the thing stumbles, a few hundred robotics researchers are watching it stumble.
Long horizon mobile manipulation is what separates demos from real deployments
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.