Google's cornhole clip skips the misses, Miki's thread doesn't

Google posted that Gemini 4 Argon watched a video of a human playing cornhole, mapped the environment, calculated whole-body joint movements, and autonomously wrote the simulation code to practice. Transferred to Apptronik's Apollo humanoid, Google says the robot sank seven straight shots in a real-world trial.

Takahiro Miki, who ran the demo, then published a nine-part breakdown of what actually happened, and it is a lot more interesting than the clip. We covered an earlier run where a model SSH'd into a robot and figured out cornhole in six tries. This is the detailed version.

What the thread shows

From a single prompt, Gemini connects to Apollo 2's stereo cameras, finds the neck joints, pans 15 degrees right onto a board 2.4 m away, and watches a human demo to pick up a slide-in strategy. It rebuilds the lab in MuJoCo (a physics simulator) and splits the board into 12 convex hulls so the 15 cm hole stays open. Solving projectile motion plus friction gives 4.55 m/s, landing 17 cm in front of the hole.

Then balance. Swinging at 5.16 rad/s creates enough reaction torque that the first stance lifts Apollo's back heel. A second iteration leaves a 1.2 cm ZMP margin, which Miki describes as near tip-over. A 52 cm wide diagonal stance gets 7.2 cm of margin and 31 percent less torque.

The misses

Throw #1 drops 85 cm short. The wrist camera slow-mo shows why. At 5.16 rad/s the 9.4 N pull beats the half-open grip of the 22-DoF Sharpa hand, which holds at 4.1 N. Fix is to lock all 22 joints to a harder clamp worth 18.5 N. Throw #2 holds with zero slip but goes 42 cm long. Throw #3 corrects depth to 2 cm. Throws #4 to #6 are three bags in a row. Throw #7 is the first airmail, a clean swish.

Then throw #8 is a rim hanger, because aiming 14.5 degrees right brings the backswing within 0.6 cm of the thigh shell. A 3D KD-tree clearance check lifts the backswing to 9.8 cm. Throw #13 is a second airmail, and #13 through #18 are six consecutive airmails, including one that knocks a hanging bag in for a six-point double sink.

So the seven straight shots are real, they just sit inside a run that starts with a bag landing most of a metre short and a grip that could not hold on. The self-correcting part is the point. Chris Paxton, amplifying the demo, summed up the reaction from the field.

Chris Paxton

@chris_j_paxton

X

Large models really are changing a lot of the assumptions about how intelligence for robotics is going to work

Takahiro Miki

@ki_ki_ki1

X

Throw #1 clears the thigh, but the bag drops 85 cm short!