live demo · about 20 seconds a run

Two robot policies, one perturbed scene. Watch them live.

Left, base π0.5. Right, ours: π0.5 trained on our targeted data sets. Both start from the same scene and run at the same time, in simulation, on one GPU. Press Run, or pick a challenge below.

Base π0.5

Ours

Robot start · level 5

The arm starts knocked out of its usual pose.

Task: pick up the black bowl between the plate and the ramekin and place it on the plate · 11 s to do it

Ready. Press Run: the first run takes about 20 seconds.

Four challenges · one click runs oneor see all 10,030 cases →

Across the whole benchmark

What you are watching

Each run is one live trial under the paper's protocol, so it can end differently from the paper's recorded result. These four cases are among those where the paper's runs separate the two models (ours succeeded in all three seeds, base π0.5 failed in every pass); each was picked from a few drawn at random for repeating live, and its card says how often it did. "Another one like this" draws from the same kind of case. Our model first rewrites the instruction with its own language model; on the language challenge, watch that line.