Hannibal52Barca/libero_baseline_pi0_pi0fast_video

LIBERO baseline rollout videos — pi0 / pi0-FAST

Rendered rollout videos from lerobot-eval runs of pi0/pi0-FAST checkpoints on the LIBERO benchmark (Franka Panda arm, hf-libero env). Each video is tagged with the checkpoint, LIBERO suite/task, task instruction, and episode success/failure.

Note: lerobot-eval only records video for the first 10 of each task's 20 evaluated episodes (batch_size=10), so per-task success rate computed from videos alone is a subsample of the full reported result, not the complete episode set.

pi0_base / pi0fast_base normalization fix

lerobot/pi0_base and lerobot/pi0fast-base ship with no normalization stats at all. LeRobot's normalizer silently passes tensors through unchanged when no stats are loaded, so their raw, unnormalized network outputs were originally sent directly to the robot as real position/orientation/gripper deltas — visible in earlier videos as wild, incoherent "flailing" rather than smooth-but-wrong motion. Their declared action/state dims (32, the cross-embodiment padded width) were also never truncated down to LIBERO's real 7/8-dim space, since the truncation step reads the dimension from the checkpoint's own config.

Fix applied: computed fresh mean/std normalization stats directly from lerobot/libero's ground-truth demonstration actions/states (273,465 frames) — independent of pi0_libero_finetuned/pi0fast-libero, so the pi0-vs-pi0-finetuned ablation isn't contaminated by borrowing from the finetuned pipeline — and patched each config's declared action/state shape from 32 (padded) to the real 7/8. Verified against pi0fast-libero's own published stats: numerically identical, confirming correctness. Videos in this dataset are post-fix: motion is now coherent (the arm reaches/moves plausibly) even though success stays ~0%, since these checkpoints have never seen LIBERO data — that part is expected, not a bug.

Structure

<model>/<suite>/task_<task_id>/episode_<n>_<success|fail>.mp4

metadata.csv at the repo root has one row per video: file_name, model, suite, task_id, task_name, episode, success, tags. task_name is the LIBERO task's natural language instruction (e.g. "pick up the black bowl next to the ramekin and place it on the plate"), from libero.libero.benchmark.

Models