Official LIBERO-Plus subset / SmolVLA · L40S

Same task.
New viewpoint.

A policy places the bowl successfully in the original scene. One official camera change alters the outcome. Replay the actual runs together, inspect the source clock and compare a light change from the same initial physical state.

Pick up the bowl and place it on the plate

Each video shows the scene camera and wrist camera. One episode per condition, selected before inference.

Original sceneSuccess · 77 actionsRecorded sample 0
Camera ViewpointsStep limit · 220 actionsRecorded sample 0
0 / 220 · 0.00 s

Loading recorded evidence…

Original episodeSuccess77 actions · 3.85 simulated seconds
Selected episodeStep limit220 actions · 11.00 simulated seconds
Initial physicsIdenticalAll qpos, qvel and source clocks match

All preselected outcomes stay visible.

Official task IDs are one-based; runtime indices are zero-based. These are three episodes, not a full benchmark.
ConditionOfficial IDOutcomeActions
Original—Success77
Camera Viewpoints609Step limit220
Light Conditions2124Success87

A concrete robustness failure to inspect. One task and one episode per condition cannot establish a model ranking or an overall success rate.

Preselected plan ↓Model & runtime ↓File fingerprints ↓

Follow the actual recorded state.

The official LIBERO-Plus wrapper changes the native camera or lighting before policy inference. Every action is newly inferred from that condition’s observations. The source records include image fingerprints, paired action-noise hashes, actual controls and simulator state.

An episode that ends earlier holds its last frame and is marked ended. The player never invents extra trajectory samples. Recorded success comes from the native task completion signal.

Original trace ↓Selected trace ↓Selected video ↓

robot-reel libero-plus verify --output docs/libero-plus --media checks hashes, clocks, controls, initial-state pairing and all video frame counts without a GPU.