LIBERO-Plus, live
Base π0.5 (left) and π0.5 trained on targeted data (right) run the same perturbed LIBERO-Plus task at the same
time, each in its own MuJoCo simulation on one GPU. Pick a perturbation, a task and a level, then run a case.
π0.5 (base)
π0.5 + targeted data
Pick a case, then press Run.
Instruction both models receive: –
Our model's reading of it: –
In the paper's evaluation pass: –
Perturbation
Task
Run
Stop
Another case
One visitor at a time: the GPU runs one session. Pressing Run opens it; it
closes after 90 s without a run.
The picker opens on the cases where the paper's runs separate the two models: base π0.5 failed in every run
and ours succeeded in every run. Across all 10,030 cases ours scores 93.2% and base 85.7%; base scores slightly
higher on lighting, background and object layout.
Each run is one live trial under the paper's protocol (seed 42, 5 flow steps, 5 actions per query). The
simulator is not bit-deterministic and both policies sample noise, so a live run can end differently from the
paper's recorded outcome for the same case. Simulation only.
Left: Physical Intelligence's public π0.5 LIBERO checkpoint (
pi05_libero ). Right: the
main model of
Find the Missing Skill, Build the Data: Targeted Data Sets for π0.5 on LIBERO-Plus .
Code and per-case results:
GitHub .
Weights:
Hugging Face .
LIBERO-Plus:
sylvestf/LIBERO-plus ; its cases and scenes run
on our server and are not redistributed by this page.