MolmoAct2 took on a five-year-old, and the gap showed
Jiafei Duan says his team entered the model in Jitendra Malik's dexterity challenge at the SPIN workshop, and the child was clearly ahead.
A robot model went head to head with a five-year-old on dexterity, and the kid won the comparison.
Jiafei Duan (@djiafei) posted that his team took part in @JitendraMalikCV's five-year-old versus robot dexterity challenge at the SPIN workshop, running MolmoAct2 against the child. His read on the result was short and not flattering to the machine.
"The gap is clear," Duan wrote, adding that "we're making progress, but robotics is far from solved." He asked whether this kind of side-by-side could be the ultimate test of robot dexterity.
we’re making progress, but robotics is far from solved
Shown versus claimed
This is a stunt, and a useful one, but it is not a benchmark. There are no published success rates, no task list and no trial counts in the post. The only description of how it went comes from the team that entered the robot, which is an unusual direction for bias to run but still a single account. Nobody outside that team has scored it.
The framing has been floating around for a couple of weeks. Brad Porter offered to race a five-year-old at CoRL before the conference, so the idea of putting a child next to a robot arm as the honest test of manipulation has been circulating in the field rather than being invented on the day.
A harder eval lands the same week
Separately, Nicolas Keller (@nicolas_keller) released what he calls the largest systematic evaluation campaign for agentic robots. His argument is that most evals treat agents as tools rather than as tool users, and that intelligent tool use is where the jump in performance comes from. He compares the shift to Homo habilis in human evolution.
Keller's thread is the opposite of the child challenge in method, scale and structure instead of one vivid comparison, and both landed the same week CoRL wrapped. Taken together they point at the same gap. The demos keep getting better, and the gap to a small child doing the same thing with their hands is still obvious enough that the team running the robot said so out loud.
most evals treat agents as tools instead of what they are: intelligent tool users
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.